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 <asm/btfixup.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 extern struct resource sparc_iomap;
  59
  60 extern unsigned long last_valid_pfn;
  61
  62 extern unsigned long page_kernel;
  63
  64 static pgd_t *srmmu_swapper_pg_dir;
  65
  66 #ifdef CONFIG_SMP
  67 #define FLUSH_BEGIN(mm)
  68 #define FLUSH_END
  69 #else
  70 #define FLUSH_BEGIN(mm) if((mm)->context != NO_CONTEXT) {
  71 #define FLUSH_END       }
  72 #endif
  73
  74 BTFIXUPDEF_CALL(void, flush_page_for_dma, unsigned long)
  75 #define flush_page_for_dma(page) BTFIXUP_CALL(flush_page_for_dma)(page)
  76
  77 int flush_page_for_dma_global = 1;
  78
  79 #ifdef CONFIG_SMP
  80 BTFIXUPDEF_CALL(void, local_flush_page_for_dma, unsigned long)
  81 #define local_flush_page_for_dma(page) BTFIXUP_CALL(local_flush_page_for_dma)(page)
  82 #endif
  83
  84 char *srmmu_name;
  85
  86 ctxd_t *srmmu_ctx_table_phys;
  87 static ctxd_t *srmmu_context_table;
  88
  89 int viking_mxcc_present;
  90 static DEFINE_SPINLOCK(srmmu_context_spinlock);
  91
  92 static int is_hypersparc;
  93
  94 /*
  95  * In general all page table modifications should use the V8 atomic
  96  * swap instruction.  This insures the mmu and the cpu are in sync
  97  * with respect to ref/mod bits in the page tables.
  98  */
  99 static inline unsigned long srmmu_swap(unsigned long *addr, unsigned long value)
 100 {
 101         __asm__ __volatile__("swap [%2], %0" : "=&r" (value) : "0" (value), "r" (addr));
 102         return value;
 103 }
 104
 105 static inline void srmmu_set_pte(pte_t *ptep, pte_t pteval)
 106 {
 107         srmmu_swap((unsigned long *)ptep, pte_val(pteval));
 108 }
 109
 110 /* The very generic SRMMU page table operations. */
 111 static inline int srmmu_device_memory(unsigned long x)
 112 {
 113         return ((x & 0xF0000000) != 0);
 114 }
 115
 116 static int srmmu_cache_pagetables;
 117
 118 /* these will be initialized in srmmu_nocache_calcsize() */
 119 static unsigned long srmmu_nocache_size;
 120 static unsigned long srmmu_nocache_end;
 121
 122 /* 1 bit <=> 256 bytes of nocache <=> 64 PTEs */
 123 #define SRMMU_NOCACHE_BITMAP_SHIFT (PAGE_SHIFT - 4)
 124
 125 /* The context table is a nocache user with the biggest alignment needs. */
 126 #define SRMMU_NOCACHE_ALIGN_MAX (sizeof(ctxd_t)*SRMMU_MAX_CONTEXTS)
 127
 128 void *srmmu_nocache_pool;
 129 void *srmmu_nocache_bitmap;
 130 static struct bit_map srmmu_nocache_map;
 131
 132 static unsigned long srmmu_pte_pfn(pte_t pte)
 133 {
 134         if (srmmu_device_memory(pte_val(pte))) {
 135                 /* Just return something that will cause
 136                  * pfn_valid() to return false.  This makes
 137                  * copy_one_pte() to just directly copy to
 138                  * PTE over.
 139                  */
 140                 return ~0UL;
 141         }
 142         return (pte_val(pte) & SRMMU_PTE_PMASK) >> (PAGE_SHIFT-4);
 143 }
 144
 145 static struct page *srmmu_pmd_page(pmd_t pmd)
 146 {
 147
 148         if (srmmu_device_memory(pmd_val(pmd)))
 149                 BUG();
 150         return pfn_to_page((pmd_val(pmd) & SRMMU_PTD_PMASK) >> (PAGE_SHIFT-4));
 151 }
 152
 153 static inline unsigned long srmmu_pgd_page(pgd_t pgd)
 154 { return srmmu_device_memory(pgd_val(pgd))?~0:(unsigned long)__nocache_va((pgd_val(pgd) & SRMMU_PTD_PMASK) << 4); }
 155
 156
 157 static inline int srmmu_pte_none(pte_t pte)
 158 { return !(pte_val(pte) & 0xFFFFFFF); }
 159
 160 static inline int srmmu_pte_present(pte_t pte)
 161 { return ((pte_val(pte) & SRMMU_ET_MASK) == SRMMU_ET_PTE); }
 162
 163 static inline void srmmu_pte_clear(pte_t *ptep)
 164 { srmmu_set_pte(ptep, __pte(0)); }
 165
 166 static inline int srmmu_pmd_none(pmd_t pmd)
 167 { return !(pmd_val(pmd) & 0xFFFFFFF); }
 168
 169 static inline int srmmu_pmd_bad(pmd_t pmd)
 170 { return (pmd_val(pmd) & SRMMU_ET_MASK) != SRMMU_ET_PTD; }
 171
 172 static inline int srmmu_pmd_present(pmd_t pmd)
 173 { return ((pmd_val(pmd) & SRMMU_ET_MASK) == SRMMU_ET_PTD); }
 174
 175 static inline void srmmu_pmd_clear(pmd_t *pmdp) {
 176         int i;
 177         for (i = 0; i < PTRS_PER_PTE/SRMMU_REAL_PTRS_PER_PTE; i++)
 178                 srmmu_set_pte((pte_t *)&pmdp->pmdv[i], __pte(0));
 179 }
 180
 181 static inline int srmmu_pgd_none(pgd_t pgd)
 182 { return !(pgd_val(pgd) & 0xFFFFFFF); }
 183
 184 static inline int srmmu_pgd_bad(pgd_t pgd)
 185 { return (pgd_val(pgd) & SRMMU_ET_MASK) != SRMMU_ET_PTD; }
 186
 187 static inline int srmmu_pgd_present(pgd_t pgd)
 188 { return ((pgd_val(pgd) & SRMMU_ET_MASK) == SRMMU_ET_PTD); }
 189
 190 static inline void srmmu_pgd_clear(pgd_t * pgdp)
 191 { srmmu_set_pte((pte_t *)pgdp, __pte(0)); }
 192
 193 static inline pte_t srmmu_pte_wrprotect(pte_t pte)
 194 { return __pte(pte_val(pte) & ~SRMMU_WRITE);}
 195
 196 static inline pte_t srmmu_pte_mkclean(pte_t pte)
 197 { return __pte(pte_val(pte) & ~SRMMU_DIRTY);}
 198
 199 static inline pte_t srmmu_pte_mkold(pte_t pte)
 200 { return __pte(pte_val(pte) & ~SRMMU_REF);}
 201
 202 static inline pte_t srmmu_pte_mkwrite(pte_t pte)
 203 { return __pte(pte_val(pte) | SRMMU_WRITE);}
 204
 205 static inline pte_t srmmu_pte_mkdirty(pte_t pte)
 206 { return __pte(pte_val(pte) | SRMMU_DIRTY);}
 207
 208 static inline pte_t srmmu_pte_mkyoung(pte_t pte)
 209 { return __pte(pte_val(pte) | SRMMU_REF);}
 210
 211 /*
 212  * Conversion functions: convert a page and protection to a page entry,
 213  * and a page entry and page directory to the page they refer to.
 214  */
 215 static pte_t srmmu_mk_pte(struct page *page, pgprot_t pgprot)
 216 { return __pte((page_to_pfn(page) << (PAGE_SHIFT-4)) | pgprot_val(pgprot)); }
 217
 218 static pte_t srmmu_mk_pte_phys(unsigned long page, pgprot_t pgprot)
 219 { return __pte(((page) >> 4) | pgprot_val(pgprot)); }
 220
 221 static pte_t srmmu_mk_pte_io(unsigned long page, pgprot_t pgprot, int space)
 222 { return __pte(((page) >> 4) | (space << 28) | pgprot_val(pgprot)); }
 223
 224 /* XXX should we hyper_flush_whole_icache here - Anton */
 225 static inline void srmmu_ctxd_set(ctxd_t *ctxp, pgd_t *pgdp)
 226 { srmmu_set_pte((pte_t *)ctxp, (SRMMU_ET_PTD | (__nocache_pa((unsigned long) pgdp) >> 4))); }
 227
 228 static inline void srmmu_pgd_set(pgd_t * pgdp, pmd_t * pmdp)
 229 { srmmu_set_pte((pte_t *)pgdp, (SRMMU_ET_PTD | (__nocache_pa((unsigned long) pmdp) >> 4))); }
 230
 231 static void srmmu_pmd_set(pmd_t *pmdp, pte_t *ptep)
 232 {
 233         unsigned long ptp;      /* Physical address, shifted right by 4 */
 234         int i;
 235
 236         ptp = __nocache_pa((unsigned long) ptep) >> 4;
 237         for (i = 0; i < PTRS_PER_PTE/SRMMU_REAL_PTRS_PER_PTE; i++) {
 238                 srmmu_set_pte((pte_t *)&pmdp->pmdv[i], SRMMU_ET_PTD | ptp);
 239                 ptp += (SRMMU_REAL_PTRS_PER_PTE*sizeof(pte_t) >> 4);
 240         }
 241 }
 242
 243 static void srmmu_pmd_populate(pmd_t *pmdp, struct page *ptep)
 244 {
 245         unsigned long ptp;      /* Physical address, shifted right by 4 */
 246         int i;
 247
 248         ptp = page_to_pfn(ptep) << (PAGE_SHIFT-4);      /* watch for overflow */
 249         for (i = 0; i < PTRS_PER_PTE/SRMMU_REAL_PTRS_PER_PTE; i++) {
 250                 srmmu_set_pte((pte_t *)&pmdp->pmdv[i], SRMMU_ET_PTD | ptp);
 251                 ptp += (SRMMU_REAL_PTRS_PER_PTE*sizeof(pte_t) >> 4);
 252         }
 253 }
 254
 255 static inline pte_t srmmu_pte_modify(pte_t pte, pgprot_t newprot)
 256 { return __pte((pte_val(pte) & SRMMU_CHG_MASK) | pgprot_val(newprot)); }
 257
 258 /* to find an entry in a top-level page table... */
 259 static inline pgd_t *srmmu_pgd_offset(struct mm_struct * mm, unsigned long address)
 260 { return mm->pgd + (address >> SRMMU_PGDIR_SHIFT); }
 261
 262 /* Find an entry in the second-level page table.. */
 263 static inline pmd_t *srmmu_pmd_offset(pgd_t * dir, unsigned long address)
 264 {
 265         return (pmd_t *) srmmu_pgd_page(*dir) +
 266             ((address >> PMD_SHIFT) & (PTRS_PER_PMD - 1));
 267 }
 268
 269 /* Find an entry in the third-level page table.. */
 270 static inline pte_t *srmmu_pte_offset(pmd_t * dir, unsigned long address)
 271 {
 272         void *pte;
 273
 274         pte = __nocache_va((dir->pmdv[0] & SRMMU_PTD_PMASK) << 4);
 275         return (pte_t *) pte +
 276             ((address >> PAGE_SHIFT) & (PTRS_PER_PTE - 1));
 277 }
 278
 279 static unsigned long srmmu_swp_type(swp_entry_t entry)
 280 {
 281         return (entry.val >> SRMMU_SWP_TYPE_SHIFT) & SRMMU_SWP_TYPE_MASK;
 282 }
 283
 284 static unsigned long srmmu_swp_offset(swp_entry_t entry)
 285 {
 286         return (entry.val >> SRMMU_SWP_OFF_SHIFT) & SRMMU_SWP_OFF_MASK;
 287 }
 288
 289 static swp_entry_t srmmu_swp_entry(unsigned long type, unsigned long offset)
 290 {
 291         return (swp_entry_t) {
 292                   (type & SRMMU_SWP_TYPE_MASK) << SRMMU_SWP_TYPE_SHIFT
 293                 | (offset & SRMMU_SWP_OFF_MASK) << SRMMU_SWP_OFF_SHIFT };
 294 }
 295
 296 /*
 297  * size: bytes to allocate in the nocache area.
 298  * align: bytes, number to align at.
 299  * Returns the virtual address of the allocated area.
 300  */
 301 static unsigned long __srmmu_get_nocache(int size, int align)
 302 {
 303         int offset;
 304
 305         if (size < SRMMU_NOCACHE_BITMAP_SHIFT) {
 306                 printk("Size 0x%x too small for nocache request\n", size);
 307                 size = SRMMU_NOCACHE_BITMAP_SHIFT;
 308         }
 309         if (size & (SRMMU_NOCACHE_BITMAP_SHIFT-1)) {
 310                 printk("Size 0x%x unaligned int nocache request\n", size);
 311                 size += SRMMU_NOCACHE_BITMAP_SHIFT-1;
 312         }
 313         BUG_ON(align > SRMMU_NOCACHE_ALIGN_MAX);
 314
 315         offset = bit_map_string_get(&srmmu_nocache_map,
 316                                         size >> SRMMU_NOCACHE_BITMAP_SHIFT,
 317                                         align >> SRMMU_NOCACHE_BITMAP_SHIFT);
 318         if (offset == -1) {
 319                 printk("srmmu: out of nocache %d: %d/%d\n",
 320                     size, (int) srmmu_nocache_size,
 321                     srmmu_nocache_map.used << SRMMU_NOCACHE_BITMAP_SHIFT);
 322                 return 0;
 323         }
 324
 325         return (SRMMU_NOCACHE_VADDR + (offset << SRMMU_NOCACHE_BITMAP_SHIFT));
 326 }
 327
 328 static unsigned long srmmu_get_nocache(int size, int align)
 329 {
 330         unsigned long tmp;
 331
 332         tmp = __srmmu_get_nocache(size, align);
 333
 334         if (tmp)
 335                 memset((void *)tmp, 0, size);
 336
 337         return tmp;
 338 }
 339
 340 static void srmmu_free_nocache(unsigned long vaddr, int size)
 341 {
 342         int offset;
 343
 344         if (vaddr < SRMMU_NOCACHE_VADDR) {
 345                 printk("Vaddr %lx is smaller than nocache base 0x%lx\n",
 346                     vaddr, (unsigned long)SRMMU_NOCACHE_VADDR);
 347                 BUG();
 348         }
 349         if (vaddr+size > srmmu_nocache_end) {
 350                 printk("Vaddr %lx is bigger than nocache end 0x%lx\n",
 351                     vaddr, srmmu_nocache_end);
 352                 BUG();
 353         }
 354         if (!is_power_of_2(size)) {
 355                 printk("Size 0x%x is not a power of 2\n", size);
 356                 BUG();
 357         }
 358         if (size < SRMMU_NOCACHE_BITMAP_SHIFT) {
 359                 printk("Size 0x%x is too small\n", size);
 360                 BUG();
 361         }
 362         if (vaddr & (size-1)) {
 363                 printk("Vaddr %lx is not aligned to size 0x%x\n", vaddr, size);
 364                 BUG();
 365         }
 366
 367         offset = (vaddr - SRMMU_NOCACHE_VADDR) >> SRMMU_NOCACHE_BITMAP_SHIFT;
 368         size = size >> SRMMU_NOCACHE_BITMAP_SHIFT;
 369
 370         bit_map_clear(&srmmu_nocache_map, offset, size);
 371 }
 372
 373 static void srmmu_early_allocate_ptable_skeleton(unsigned long start,
 374                                                  unsigned long end);
 375
 376 extern unsigned long probe_memory(void);        /* in fault.c */
 377
 378 /*
 379  * Reserve nocache dynamically proportionally to the amount of
 380  * system RAM. -- Tomas Szepe <szepe@pinerecords.com>, June 2002
 381  */
 382 static void srmmu_nocache_calcsize(void)
 383 {
 384         unsigned long sysmemavail = probe_memory() / 1024;
 385         int srmmu_nocache_npages;
 386
 387         srmmu_nocache_npages =
 388                 sysmemavail / SRMMU_NOCACHE_ALCRATIO / 1024 * 256;
 389
 390  /* P3 XXX The 4x overuse: corroborated by /proc/meminfo. */
 391         // if (srmmu_nocache_npages < 256) srmmu_nocache_npages = 256;
 392         if (srmmu_nocache_npages < SRMMU_MIN_NOCACHE_PAGES)
 393                 srmmu_nocache_npages = SRMMU_MIN_NOCACHE_PAGES;
 394
 395         /* anything above 1280 blows up */
 396         if (srmmu_nocache_npages > SRMMU_MAX_NOCACHE_PAGES)
 397                 srmmu_nocache_npages = SRMMU_MAX_NOCACHE_PAGES;
 398
 399         srmmu_nocache_size = srmmu_nocache_npages * PAGE_SIZE;
 400         srmmu_nocache_end = SRMMU_NOCACHE_VADDR + srmmu_nocache_size;
 401 }
 402
 403 static void __init srmmu_nocache_init(void)
 404 {
 405         unsigned int bitmap_bits;
 406         pgd_t *pgd;
 407         pmd_t *pmd;
 408         pte_t *pte;
 409         unsigned long paddr, vaddr;
 410         unsigned long pteval;
 411
 412         bitmap_bits = srmmu_nocache_size >> SRMMU_NOCACHE_BITMAP_SHIFT;
 413
 414         srmmu_nocache_pool = __alloc_bootmem(srmmu_nocache_size,
 415                 SRMMU_NOCACHE_ALIGN_MAX, 0UL);
 416         memset(srmmu_nocache_pool, 0, srmmu_nocache_size);
 417
 418         srmmu_nocache_bitmap = __alloc_bootmem(bitmap_bits >> 3, SMP_CACHE_BYTES, 0UL);
 419         bit_map_init(&srmmu_nocache_map, srmmu_nocache_bitmap, bitmap_bits);
 420
 421         srmmu_swapper_pg_dir = (pgd_t *)__srmmu_get_nocache(SRMMU_PGD_TABLE_SIZE, SRMMU_PGD_TABLE_SIZE);
 422         memset(__nocache_fix(srmmu_swapper_pg_dir), 0, SRMMU_PGD_TABLE_SIZE);
 423         init_mm.pgd = srmmu_swapper_pg_dir;
 424
 425         srmmu_early_allocate_ptable_skeleton(SRMMU_NOCACHE_VADDR, srmmu_nocache_end);
 426
 427         paddr = __pa((unsigned long)srmmu_nocache_pool);
 428         vaddr = SRMMU_NOCACHE_VADDR;
 429
 430         while (vaddr < srmmu_nocache_end) {
 431                 pgd = pgd_offset_k(vaddr);
 432                 pmd = srmmu_pmd_offset(__nocache_fix(pgd), vaddr);
 433                 pte = srmmu_pte_offset(__nocache_fix(pmd), vaddr);
 434
 435                 pteval = ((paddr >> 4) | SRMMU_ET_PTE | SRMMU_PRIV);
 436
 437                 if (srmmu_cache_pagetables)
 438                         pteval |= SRMMU_CACHE;
 439
 440                 srmmu_set_pte(__nocache_fix(pte), __pte(pteval));
 441
 442                 vaddr += PAGE_SIZE;
 443                 paddr += PAGE_SIZE;
 444         }
 445
 446         flush_cache_all();
 447         flush_tlb_all();
 448 }
 449
 450 static inline pgd_t *srmmu_get_pgd_fast(void)
 451 {
 452         pgd_t *pgd = NULL;
 453
 454         pgd = (pgd_t *)__srmmu_get_nocache(SRMMU_PGD_TABLE_SIZE, SRMMU_PGD_TABLE_SIZE);
 455         if (pgd) {
 456                 pgd_t *init = pgd_offset_k(0);
 457                 memset(pgd, 0, USER_PTRS_PER_PGD * sizeof(pgd_t));
 458                 memcpy(pgd + USER_PTRS_PER_PGD, init + USER_PTRS_PER_PGD,
 459                                                 (PTRS_PER_PGD - USER_PTRS_PER_PGD) * sizeof(pgd_t));
 460         }
 461
 462         return pgd;
 463 }
 464
 465 static void srmmu_free_pgd_fast(pgd_t *pgd)
 466 {
 467         srmmu_free_nocache((unsigned long)pgd, SRMMU_PGD_TABLE_SIZE);
 468 }
 469
 470 static pmd_t *srmmu_pmd_alloc_one(struct mm_struct *mm, unsigned long address)
 471 {
 472         return (pmd_t *)srmmu_get_nocache(SRMMU_PMD_TABLE_SIZE, SRMMU_PMD_TABLE_SIZE);
 473 }
 474
 475 static void srmmu_pmd_free(pmd_t * pmd)
 476 {
 477         srmmu_free_nocache((unsigned long)pmd, SRMMU_PMD_TABLE_SIZE);
 478 }
 479
 480 /*
 481  * Hardware needs alignment to 256 only, but we align to whole page size
 482  * to reduce fragmentation problems due to the buddy principle.
 483  * XXX Provide actual fragmentation statistics in /proc.
 484  *
 485  * Alignments up to the page size are the same for physical and virtual
 486  * addresses of the nocache area.
 487  */
 488 static pte_t *
 489 srmmu_pte_alloc_one_kernel(struct mm_struct *mm, unsigned long address)
 490 {
 491         return (pte_t *)srmmu_get_nocache(PTE_SIZE, PTE_SIZE);
 492 }
 493
 494 static pgtable_t
 495 srmmu_pte_alloc_one(struct mm_struct *mm, unsigned long address)
 496 {
 497         unsigned long pte;
 498         struct page *page;
 499
 500         if ((pte = (unsigned long)srmmu_pte_alloc_one_kernel(mm, address)) == 0)
 501                 return NULL;
 502         page = pfn_to_page( __nocache_pa(pte) >> PAGE_SHIFT );
 503         pgtable_page_ctor(page);
 504         return page;
 505 }
 506
 507 static void srmmu_free_pte_fast(pte_t *pte)
 508 {
 509         srmmu_free_nocache((unsigned long)pte, PTE_SIZE);
 510 }
 511
 512 static void srmmu_pte_free(pgtable_t pte)
 513 {
 514         unsigned long p;
 515
 516         pgtable_page_dtor(pte);
 517         p = (unsigned long)page_address(pte);   /* Cached address (for test) */
 518         if (p == 0)
 519                 BUG();
 520         p = page_to_pfn(pte) << PAGE_SHIFT;     /* Physical address */
 521         p = (unsigned long) __nocache_va(p);    /* Nocached virtual */
 522         srmmu_free_nocache(p, PTE_SIZE);
 523 }
 524
 525 /*
 526  */
 527 static inline void alloc_context(struct mm_struct *old_mm, struct mm_struct *mm)
 528 {
 529         struct ctx_list *ctxp;
 530
 531         ctxp = ctx_free.next;
 532         if(ctxp != &ctx_free) {
 533                 remove_from_ctx_list(ctxp);
 534                 add_to_used_ctxlist(ctxp);
 535                 mm->context = ctxp->ctx_number;
 536                 ctxp->ctx_mm = mm;
 537                 return;
 538         }
 539         ctxp = ctx_used.next;
 540         if(ctxp->ctx_mm == old_mm)
 541                 ctxp = ctxp->next;
 542         if(ctxp == &ctx_used)
 543                 panic("out of mmu contexts");
 544         flush_cache_mm(ctxp->ctx_mm);
 545         flush_tlb_mm(ctxp->ctx_mm);
 546         remove_from_ctx_list(ctxp);
 547         add_to_used_ctxlist(ctxp);
 548         ctxp->ctx_mm->context = NO_CONTEXT;
 549         ctxp->ctx_mm = mm;
 550         mm->context = ctxp->ctx_number;
 551 }
 552
 553 static inline void free_context(int context)
 554 {
 555         struct ctx_list *ctx_old;
 556
 557         ctx_old = ctx_list_pool + context;
 558         remove_from_ctx_list(ctx_old);
 559         add_to_free_ctxlist(ctx_old);
 560 }
 561
 562
 563 static void srmmu_switch_mm(struct mm_struct *old_mm, struct mm_struct *mm,
 564     struct task_struct *tsk, int cpu)
 565 {
 566         if(mm->context == NO_CONTEXT) {
 567                 spin_lock(&srmmu_context_spinlock);
 568                 alloc_context(old_mm, mm);
 569                 spin_unlock(&srmmu_context_spinlock);
 570                 srmmu_ctxd_set(&srmmu_context_table[mm->context], mm->pgd);
 571         }
 572
 573         if (sparc_cpu_model == sparc_leon)
 574                 leon_switch_mm();
 575
 576         if (is_hypersparc)
 577                 hyper_flush_whole_icache();
 578
 579         srmmu_set_context(mm->context);
 580 }
 581
 582 /* Low level IO area allocation on the SRMMU. */
 583 static inline void srmmu_mapioaddr(unsigned long physaddr,
 584     unsigned long virt_addr, int bus_type)
 585 {
 586         pgd_t *pgdp;
 587         pmd_t *pmdp;
 588         pte_t *ptep;
 589         unsigned long tmp;
 590
 591         physaddr &= PAGE_MASK;
 592         pgdp = pgd_offset_k(virt_addr);
 593         pmdp = srmmu_pmd_offset(pgdp, virt_addr);
 594         ptep = srmmu_pte_offset(pmdp, virt_addr);
 595         tmp = (physaddr >> 4) | SRMMU_ET_PTE;
 596
 597         /*
 598          * I need to test whether this is consistent over all
 599          * sun4m's.  The bus_type represents the upper 4 bits of
 600          * 36-bit physical address on the I/O space lines...
 601          */
 602         tmp |= (bus_type << 28);
 603         tmp |= SRMMU_PRIV;
 604         __flush_page_to_ram(virt_addr);
 605         srmmu_set_pte(ptep, __pte(tmp));
 606 }
 607
 608 static void srmmu_mapiorange(unsigned int bus, unsigned long xpa,
 609     unsigned long xva, unsigned int len)
 610 {
 611         while (len != 0) {
 612                 len -= PAGE_SIZE;
 613                 srmmu_mapioaddr(xpa, xva, bus);
 614                 xva += PAGE_SIZE;
 615                 xpa += PAGE_SIZE;
 616         }
 617         flush_tlb_all();
 618 }
 619
 620 static inline void srmmu_unmapioaddr(unsigned long virt_addr)
 621 {
 622         pgd_t *pgdp;
 623         pmd_t *pmdp;
 624         pte_t *ptep;
 625
 626         pgdp = pgd_offset_k(virt_addr);
 627         pmdp = srmmu_pmd_offset(pgdp, virt_addr);
 628         ptep = srmmu_pte_offset(pmdp, virt_addr);
 629
 630         /* No need to flush uncacheable page. */
 631         srmmu_pte_clear(ptep);
 632 }
 633
 634 static void srmmu_unmapiorange(unsigned long virt_addr, unsigned int len)
 635 {
 636         while (len != 0) {
 637                 len -= PAGE_SIZE;
 638                 srmmu_unmapioaddr(virt_addr);
 639                 virt_addr += PAGE_SIZE;
 640         }
 641         flush_tlb_all();
 642 }
 643
 644 /*
 645  * On the SRMMU we do not have the problems with limited tlb entries
 646  * for mapping kernel pages, so we just take things from the free page
 647  * pool.  As a side effect we are putting a little too much pressure
 648  * on the gfp() subsystem.  This setup also makes the logic of the
 649  * iommu mapping code a lot easier as we can transparently handle
 650  * mappings on the kernel stack without any special code as we did
 651  * need on the sun4c.
 652  */
 653 static struct thread_info *srmmu_alloc_thread_info(void)
 654 {
 655         struct thread_info *ret;
 656
 657         ret = (struct thread_info *)__get_free_pages(GFP_KERNEL,
 658                                                      THREAD_INFO_ORDER);
 659 #ifdef CONFIG_DEBUG_STACK_USAGE
 660         if (ret)
 661                 memset(ret, 0, PAGE_SIZE << THREAD_INFO_ORDER);
 662 #endif /* DEBUG_STACK_USAGE */
 663
 664         return ret;
 665 }
 666
 667 static void srmmu_free_thread_info(struct thread_info *ti)
 668 {
 669         free_pages((unsigned long)ti, THREAD_INFO_ORDER);
 670 }
 671
 672 /* tsunami.S */
 673 extern void tsunami_flush_cache_all(void);
 674 extern void tsunami_flush_cache_mm(struct mm_struct *mm);
 675 extern void tsunami_flush_cache_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
 676 extern void tsunami_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
 677 extern void tsunami_flush_page_to_ram(unsigned long page);
 678 extern void tsunami_flush_page_for_dma(unsigned long page);
 679 extern void tsunami_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr);
 680 extern void tsunami_flush_tlb_all(void);
 681 extern void tsunami_flush_tlb_mm(struct mm_struct *mm);
 682 extern void tsunami_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
 683 extern void tsunami_flush_tlb_page(struct vm_area_struct *vma, unsigned long page);
 684 extern void tsunami_setup_blockops(void);
 685
 686 /*
 687  * Workaround, until we find what's going on with Swift. When low on memory,
 688  * it sometimes loops in fault/handle_mm_fault incl. flush_tlb_page to find
 689  * out it is already in page tables/ fault again on the same instruction.
 690  * I really don't understand it, have checked it and contexts
 691  * are right, flush_tlb_all is done as well, and it faults again...
 692  * Strange. -jj
 693  *
 694  * The following code is a deadwood that may be necessary when
 695  * we start to make precise page flushes again. --zaitcev
 696  */
 697 static void swift_update_mmu_cache(struct vm_area_struct * vma, unsigned long address, pte_t *ptep)
 698 {
 699 #if 0
 700         static unsigned long last;
 701         unsigned int val;
 702         /* unsigned int n; */
 703
 704         if (address == last) {
 705                 val = srmmu_hwprobe(address);
 706                 if (val != 0 && pte_val(*ptep) != val) {
 707                         printk("swift_update_mmu_cache: "
 708                             "addr %lx put %08x probed %08x from %p\n",
 709                             address, pte_val(*ptep), val,
 710                             __builtin_return_address(0));
 711                         srmmu_flush_whole_tlb();
 712                 }
 713         }
 714         last = address;
 715 #endif
 716 }
 717
 718 /* swift.S */
 719 extern void swift_flush_cache_all(void);
 720 extern void swift_flush_cache_mm(struct mm_struct *mm);
 721 extern void swift_flush_cache_range(struct vm_area_struct *vma,
 722                                     unsigned long start, unsigned long end);
 723 extern void swift_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
 724 extern void swift_flush_page_to_ram(unsigned long page);
 725 extern void swift_flush_page_for_dma(unsigned long page);
 726 extern void swift_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr);
 727 extern void swift_flush_tlb_all(void);
 728 extern void swift_flush_tlb_mm(struct mm_struct *mm);
 729 extern void swift_flush_tlb_range(struct vm_area_struct *vma,
 730                                   unsigned long start, unsigned long end);
 731 extern void swift_flush_tlb_page(struct vm_area_struct *vma, unsigned long page);
 732
 733 #if 0  /* P3: deadwood to debug precise flushes on Swift. */
 734 void swift_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
 735 {
 736         int cctx, ctx1;
 737
 738         page &= PAGE_MASK;
 739         if ((ctx1 = vma->vm_mm->context) != -1) {
 740                 cctx = srmmu_get_context();
 741 /* Is context # ever different from current context? P3 */
 742                 if (cctx != ctx1) {
 743                         printk("flush ctx %02x curr %02x\n", ctx1, cctx);
 744                         srmmu_set_context(ctx1);
 745                         swift_flush_page(page);
 746                         __asm__ __volatile__("sta %%g0, [%0] %1\n\t" : :
 747                                         "r" (page), "i" (ASI_M_FLUSH_PROBE));
 748                         srmmu_set_context(cctx);
 749                 } else {
 750                          /* Rm. prot. bits from virt. c. */
 751                         /* swift_flush_cache_all(); */
 752                         /* swift_flush_cache_page(vma, page); */
 753                         swift_flush_page(page);
 754
 755                         __asm__ __volatile__("sta %%g0, [%0] %1\n\t" : :
 756                                 "r" (page), "i" (ASI_M_FLUSH_PROBE));
 757                         /* same as above: srmmu_flush_tlb_page() */
 758                 }
 759         }
 760 }
 761 #endif
 762
 763 /*
 764  * The following are all MBUS based SRMMU modules, and therefore could
 765  * be found in a multiprocessor configuration.  On the whole, these
 766  * chips seems to be much more touchy about DVMA and page tables
 767  * with respect to cache coherency.
 768  */
 769
 770 /* Cypress flushes. */
 771 static void cypress_flush_cache_all(void)
 772 {
 773         volatile unsigned long cypress_sucks;
 774         unsigned long faddr, tagval;
 775
 776         flush_user_windows();
 777         for(faddr = 0; faddr < 0x10000; faddr += 0x20) {
 778                 __asm__ __volatile__("lda [%1 + %2] %3, %0\n\t" :
 779                                      "=r" (tagval) :
 780                                      "r" (faddr), "r" (0x40000),
 781                                      "i" (ASI_M_DATAC_TAG));
 782
 783                 /* If modified and valid, kick it. */
 784                 if((tagval & 0x60) == 0x60)
 785                         cypress_sucks = *(unsigned long *)(0xf0020000 + faddr);
 786         }
 787 }
 788
 789 static void cypress_flush_cache_mm(struct mm_struct *mm)
 790 {
 791         register unsigned long a, b, c, d, e, f, g;
 792         unsigned long flags, faddr;
 793         int octx;
 794
 795         FLUSH_BEGIN(mm)
 796         flush_user_windows();
 797         local_irq_save(flags);
 798         octx = srmmu_get_context();
 799         srmmu_set_context(mm->context);
 800         a = 0x20; b = 0x40; c = 0x60;
 801         d = 0x80; e = 0xa0; f = 0xc0; g = 0xe0;
 802
 803         faddr = (0x10000 - 0x100);
 804         goto inside;
 805         do {
 806                 faddr -= 0x100;
 807         inside:
 808                 __asm__ __volatile__("sta %%g0, [%0] %1\n\t"
 809                                      "sta %%g0, [%0 + %2] %1\n\t"
 810                                      "sta %%g0, [%0 + %3] %1\n\t"
 811                                      "sta %%g0, [%0 + %4] %1\n\t"
 812                                      "sta %%g0, [%0 + %5] %1\n\t"
 813                                      "sta %%g0, [%0 + %6] %1\n\t"
 814                                      "sta %%g0, [%0 + %7] %1\n\t"
 815                                      "sta %%g0, [%0 + %8] %1\n\t" : :
 816                                      "r" (faddr), "i" (ASI_M_FLUSH_CTX),
 817                                      "r" (a), "r" (b), "r" (c), "r" (d),
 818                                      "r" (e), "r" (f), "r" (g));
 819         } while(faddr);
 820         srmmu_set_context(octx);
 821         local_irq_restore(flags);
 822         FLUSH_END
 823 }
 824
 825 static void cypress_flush_cache_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
 826 {
 827         struct mm_struct *mm = vma->vm_mm;
 828         register unsigned long a, b, c, d, e, f, g;
 829         unsigned long flags, faddr;
 830         int octx;
 831
 832         FLUSH_BEGIN(mm)
 833         flush_user_windows();
 834         local_irq_save(flags);
 835         octx = srmmu_get_context();
 836         srmmu_set_context(mm->context);
 837         a = 0x20; b = 0x40; c = 0x60;
 838         d = 0x80; e = 0xa0; f = 0xc0; g = 0xe0;
 839
 840         start &= SRMMU_REAL_PMD_MASK;
 841         while(start < end) {
 842                 faddr = (start + (0x10000 - 0x100));
 843                 goto inside;
 844                 do {
 845                         faddr -= 0x100;
 846                 inside:
 847                         __asm__ __volatile__("sta %%g0, [%0] %1\n\t"
 848                                              "sta %%g0, [%0 + %2] %1\n\t"
 849                                              "sta %%g0, [%0 + %3] %1\n\t"
 850                                              "sta %%g0, [%0 + %4] %1\n\t"
 851                                              "sta %%g0, [%0 + %5] %1\n\t"
 852                                              "sta %%g0, [%0 + %6] %1\n\t"
 853                                              "sta %%g0, [%0 + %7] %1\n\t"
 854                                              "sta %%g0, [%0 + %8] %1\n\t" : :
 855                                              "r" (faddr),
 856                                              "i" (ASI_M_FLUSH_SEG),
 857                                              "r" (a), "r" (b), "r" (c), "r" (d),
 858                                              "r" (e), "r" (f), "r" (g));
 859                 } while (faddr != start);
 860                 start += SRMMU_REAL_PMD_SIZE;
 861         }
 862         srmmu_set_context(octx);
 863         local_irq_restore(flags);
 864         FLUSH_END
 865 }
 866
 867 static void cypress_flush_cache_page(struct vm_area_struct *vma, unsigned long page)
 868 {
 869         register unsigned long a, b, c, d, e, f, g;
 870         struct mm_struct *mm = vma->vm_mm;
 871         unsigned long flags, line;
 872         int octx;
 873
 874         FLUSH_BEGIN(mm)
 875         flush_user_windows();
 876         local_irq_save(flags);
 877         octx = srmmu_get_context();
 878         srmmu_set_context(mm->context);
 879         a = 0x20; b = 0x40; c = 0x60;
 880         d = 0x80; e = 0xa0; f = 0xc0; g = 0xe0;
 881
 882         page &= PAGE_MASK;
 883         line = (page + PAGE_SIZE) - 0x100;
 884         goto inside;
 885         do {
 886                 line -= 0x100;
 887         inside:
 888                         __asm__ __volatile__("sta %%g0, [%0] %1\n\t"
 889                                              "sta %%g0, [%0 + %2] %1\n\t"
 890                                              "sta %%g0, [%0 + %3] %1\n\t"
 891                                              "sta %%g0, [%0 + %4] %1\n\t"
 892                                              "sta %%g0, [%0 + %5] %1\n\t"
 893                                              "sta %%g0, [%0 + %6] %1\n\t"
 894                                              "sta %%g0, [%0 + %7] %1\n\t"
 895                                              "sta %%g0, [%0 + %8] %1\n\t" : :
 896                                              "r" (line),
 897                                              "i" (ASI_M_FLUSH_PAGE),
 898                                              "r" (a), "r" (b), "r" (c), "r" (d),
 899                                              "r" (e), "r" (f), "r" (g));
 900         } while(line != page);
 901         srmmu_set_context(octx);
 902         local_irq_restore(flags);
 903         FLUSH_END
 904 }
 905
 906 /* Cypress is copy-back, at least that is how we configure it. */
 907 static void cypress_flush_page_to_ram(unsigned long page)
 908 {
 909         register unsigned long a, b, c, d, e, f, g;
 910         unsigned long line;
 911
 912         a = 0x20; b = 0x40; c = 0x60; d = 0x80; e = 0xa0; f = 0xc0; g = 0xe0;
 913         page &= PAGE_MASK;
 914         line = (page + PAGE_SIZE) - 0x100;
 915         goto inside;
 916         do {
 917                 line -= 0x100;
 918         inside:
 919                 __asm__ __volatile__("sta %%g0, [%0] %1\n\t"
 920                                      "sta %%g0, [%0 + %2] %1\n\t"
 921                                      "sta %%g0, [%0 + %3] %1\n\t"
 922                                      "sta %%g0, [%0 + %4] %1\n\t"
 923                                      "sta %%g0, [%0 + %5] %1\n\t"
 924                                      "sta %%g0, [%0 + %6] %1\n\t"
 925                                      "sta %%g0, [%0 + %7] %1\n\t"
 926                                      "sta %%g0, [%0 + %8] %1\n\t" : :
 927                                      "r" (line),
 928                                      "i" (ASI_M_FLUSH_PAGE),
 929                                      "r" (a), "r" (b), "r" (c), "r" (d),
 930                                      "r" (e), "r" (f), "r" (g));
 931         } while(line != page);
 932 }
 933
 934 /* Cypress is also IO cache coherent. */
 935 static void cypress_flush_page_for_dma(unsigned long page)
 936 {
 937 }
 938
 939 /* Cypress has unified L2 VIPT, from which both instructions and data
 940  * are stored.  It does not have an onboard icache of any sort, therefore
 941  * no flush is necessary.
 942  */
 943 static void cypress_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr)
 944 {
 945 }
 946
 947 static void cypress_flush_tlb_all(void)
 948 {
 949         srmmu_flush_whole_tlb();
 950 }
 951
 952 static void cypress_flush_tlb_mm(struct mm_struct *mm)
 953 {
 954         FLUSH_BEGIN(mm)
 955         __asm__ __volatile__(
 956         "lda    [%0] %3, %%g5\n\t"
 957         "sta    %2, [%0] %3\n\t"
 958         "sta    %%g0, [%1] %4\n\t"
 959         "sta    %%g5, [%0] %3\n"
 960         : /* no outputs */
 961         : "r" (SRMMU_CTX_REG), "r" (0x300), "r" (mm->context),
 962           "i" (ASI_M_MMUREGS), "i" (ASI_M_FLUSH_PROBE)
 963         : "g5");
 964         FLUSH_END
 965 }
 966
 967 static void cypress_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
 968 {
 969         struct mm_struct *mm = vma->vm_mm;
 970         unsigned long size;
 971
 972         FLUSH_BEGIN(mm)
 973         start &= SRMMU_PGDIR_MASK;
 974         size = SRMMU_PGDIR_ALIGN(end) - start;
 975         __asm__ __volatile__(
 976                 "lda    [%0] %5, %%g5\n\t"
 977                 "sta    %1, [%0] %5\n"
 978                 "1:\n\t"
 979                 "subcc  %3, %4, %3\n\t"
 980                 "bne    1b\n\t"
 981                 " sta   %%g0, [%2 + %3] %6\n\t"
 982                 "sta    %%g5, [%0] %5\n"
 983         : /* no outputs */
 984         : "r" (SRMMU_CTX_REG), "r" (mm->context), "r" (start | 0x200),
 985           "r" (size), "r" (SRMMU_PGDIR_SIZE), "i" (ASI_M_MMUREGS),
 986           "i" (ASI_M_FLUSH_PROBE)
 987         : "g5", "cc");
 988         FLUSH_END
 989 }
 990
 991 static void cypress_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
 992 {
 993         struct mm_struct *mm = vma->vm_mm;
 994
 995         FLUSH_BEGIN(mm)
 996         __asm__ __volatile__(
 997         "lda    [%0] %3, %%g5\n\t"
 998         "sta    %1, [%0] %3\n\t"
 999         "sta    %%g0, [%2] %4\n\t"
1000         "sta    %%g5, [%0] %3\n"
1001         : /* no outputs */
1002         : "r" (SRMMU_CTX_REG), "r" (mm->context), "r" (page & PAGE_MASK),
1003           "i" (ASI_M_MMUREGS), "i" (ASI_M_FLUSH_PROBE)
1004         : "g5");
1005         FLUSH_END
1006 }
1007
1008 /* viking.S */
1009 extern void viking_flush_cache_all(void);
1010 extern void viking_flush_cache_mm(struct mm_struct *mm);
1011 extern void viking_flush_cache_range(struct vm_area_struct *vma, unsigned long start,
1012                                      unsigned long end);
1013 extern void viking_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
1014 extern void viking_flush_page_to_ram(unsigned long page);
1015 extern void viking_flush_page_for_dma(unsigned long page);
1016 extern void viking_flush_sig_insns(struct mm_struct *mm, unsigned long addr);
1017 extern void viking_flush_page(unsigned long page);
1018 extern void viking_mxcc_flush_page(unsigned long page);
1019 extern void viking_flush_tlb_all(void);
1020 extern void viking_flush_tlb_mm(struct mm_struct *mm);
1021 extern void viking_flush_tlb_range(struct vm_area_struct *vma, unsigned long start,
1022                                    unsigned long end);
1023 extern void viking_flush_tlb_page(struct vm_area_struct *vma,
1024                                   unsigned long page);
1025 extern void sun4dsmp_flush_tlb_all(void);
1026 extern void sun4dsmp_flush_tlb_mm(struct mm_struct *mm);
1027 extern void sun4dsmp_flush_tlb_range(struct vm_area_struct *vma, unsigned long start,
1028                                    unsigned long end);
1029 extern void sun4dsmp_flush_tlb_page(struct vm_area_struct *vma,
1030                                   unsigned long page);
1031
1032 /* hypersparc.S */
1033 extern void hypersparc_flush_cache_all(void);
1034 extern void hypersparc_flush_cache_mm(struct mm_struct *mm);
1035 extern void hypersparc_flush_cache_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
1036 extern void hypersparc_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
1037 extern void hypersparc_flush_page_to_ram(unsigned long page);
1038 extern void hypersparc_flush_page_for_dma(unsigned long page);
1039 extern void hypersparc_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr);
1040 extern void hypersparc_flush_tlb_all(void);
1041 extern void hypersparc_flush_tlb_mm(struct mm_struct *mm);
1042 extern void hypersparc_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
1043 extern void hypersparc_flush_tlb_page(struct vm_area_struct *vma, unsigned long page);
1044 extern void hypersparc_setup_blockops(void);
1045
1046 /*
1047  * NOTE: All of this startup code assumes the low 16mb (approx.) of
1048  *       kernel mappings are done with one single contiguous chunk of
1049  *       ram.  On small ram machines (classics mainly) we only get
1050  *       around 8mb mapped for us.
1051  */
1052
1053 static void __init early_pgtable_allocfail(char *type)
1054 {
1055         prom_printf("inherit_prom_mappings: Cannot alloc kernel %s.\n", type);
1056         prom_halt();
1057 }
1058
1059 static void __init srmmu_early_allocate_ptable_skeleton(unsigned long start,
1060                                                         unsigned long end)
1061 {
1062         pgd_t *pgdp;
1063         pmd_t *pmdp;
1064         pte_t *ptep;
1065
1066         while(start < end) {
1067                 pgdp = pgd_offset_k(start);
1068                 if(srmmu_pgd_none(*(pgd_t *)__nocache_fix(pgdp))) {
1069                         pmdp = (pmd_t *) __srmmu_get_nocache(
1070                             SRMMU_PMD_TABLE_SIZE, SRMMU_PMD_TABLE_SIZE);
1071                         if (pmdp == NULL)
1072                                 early_pgtable_allocfail("pmd");
1073                         memset(__nocache_fix(pmdp), 0, SRMMU_PMD_TABLE_SIZE);
1074                         srmmu_pgd_set(__nocache_fix(pgdp), pmdp);
1075                 }
1076                 pmdp = srmmu_pmd_offset(__nocache_fix(pgdp), start);
1077                 if(srmmu_pmd_none(*(pmd_t *)__nocache_fix(pmdp))) {
1078                         ptep = (pte_t *)__srmmu_get_nocache(PTE_SIZE, PTE_SIZE);
1079                         if (ptep == NULL)
1080                                 early_pgtable_allocfail("pte");
1081                         memset(__nocache_fix(ptep), 0, PTE_SIZE);
1082                         srmmu_pmd_set(__nocache_fix(pmdp), ptep);
1083                 }
1084                 if (start > (0xffffffffUL - PMD_SIZE))
1085                         break;
1086                 start = (start + PMD_SIZE) & PMD_MASK;
1087         }
1088 }
1089
1090 static void __init srmmu_allocate_ptable_skeleton(unsigned long start,
1091                                                   unsigned long end)
1092 {
1093         pgd_t *pgdp;
1094         pmd_t *pmdp;
1095         pte_t *ptep;
1096
1097         while(start < end) {
1098                 pgdp = pgd_offset_k(start);
1099                 if(srmmu_pgd_none(*pgdp)) {
1100                         pmdp = (pmd_t *)__srmmu_get_nocache(SRMMU_PMD_TABLE_SIZE, SRMMU_PMD_TABLE_SIZE);
1101                         if (pmdp == NULL)
1102                                 early_pgtable_allocfail("pmd");
1103                         memset(pmdp, 0, SRMMU_PMD_TABLE_SIZE);
1104                         srmmu_pgd_set(pgdp, pmdp);
1105                 }
1106                 pmdp = srmmu_pmd_offset(pgdp, start);
1107                 if(srmmu_pmd_none(*pmdp)) {
1108                         ptep = (pte_t *) __srmmu_get_nocache(PTE_SIZE,
1109                                                              PTE_SIZE);
1110                         if (ptep == NULL)
1111                                 early_pgtable_allocfail("pte");
1112                         memset(ptep, 0, PTE_SIZE);
1113                         srmmu_pmd_set(pmdp, ptep);
1114                 }
1115                 if (start > (0xffffffffUL - PMD_SIZE))
1116                         break;
1117                 start = (start + PMD_SIZE) & PMD_MASK;
1118         }
1119 }
1120
1121 /*
1122  * This is much cleaner than poking around physical address space
1123  * looking at the prom's page table directly which is what most
1124  * other OS's do.  Yuck... this is much better.
1125  */
1126 static void __init srmmu_inherit_prom_mappings(unsigned long start,
1127                                                unsigned long end)
1128 {
1129         pgd_t *pgdp;
1130         pmd_t *pmdp;
1131         pte_t *ptep;
1132         int what = 0; /* 0 = normal-pte, 1 = pmd-level pte, 2 = pgd-level pte */
1133         unsigned long prompte;
1134
1135         while(start <= end) {
1136                 if (start == 0)
1137                         break; /* probably wrap around */
1138                 if(start == 0xfef00000)
1139                         start = KADB_DEBUGGER_BEGVM;
1140                 if(!(prompte = srmmu_hwprobe(start))) {
1141                         start += PAGE_SIZE;
1142                         continue;
1143                 }
1144
1145                 /* A red snapper, see what it really is. */
1146                 what = 0;
1147
1148                 if(!(start & ~(SRMMU_REAL_PMD_MASK))) {
1149                         if(srmmu_hwprobe((start-PAGE_SIZE) + SRMMU_REAL_PMD_SIZE) == prompte)
1150                                 what = 1;
1151                 }
1152
1153                 if(!(start & ~(SRMMU_PGDIR_MASK))) {
1154                         if(srmmu_hwprobe((start-PAGE_SIZE) + SRMMU_PGDIR_SIZE) ==
1155                            prompte)
1156                                 what = 2;
1157                 }
1158
1159                 pgdp = pgd_offset_k(start);
1160                 if(what == 2) {
1161                         *(pgd_t *)__nocache_fix(pgdp) = __pgd(prompte);
1162                         start += SRMMU_PGDIR_SIZE;
1163                         continue;
1164                 }
1165                 if(srmmu_pgd_none(*(pgd_t *)__nocache_fix(pgdp))) {
1166                         pmdp = (pmd_t *)__srmmu_get_nocache(SRMMU_PMD_TABLE_SIZE, SRMMU_PMD_TABLE_SIZE);
1167                         if (pmdp == NULL)
1168                                 early_pgtable_allocfail("pmd");
1169                         memset(__nocache_fix(pmdp), 0, SRMMU_PMD_TABLE_SIZE);
1170                         srmmu_pgd_set(__nocache_fix(pgdp), pmdp);
1171                 }
1172                 pmdp = srmmu_pmd_offset(__nocache_fix(pgdp), start);
1173                 if(srmmu_pmd_none(*(pmd_t *)__nocache_fix(pmdp))) {
1174                         ptep = (pte_t *) __srmmu_get_nocache(PTE_SIZE,
1175                                                              PTE_SIZE);
1176                         if (ptep == NULL)
1177                                 early_pgtable_allocfail("pte");
1178                         memset(__nocache_fix(ptep), 0, PTE_SIZE);
1179                         srmmu_pmd_set(__nocache_fix(pmdp), ptep);
1180                 }
1181                 if(what == 1) {
1182                         /*
1183                          * We bend the rule where all 16 PTPs in a pmd_t point
1184                          * inside the same PTE page, and we leak a perfectly
1185                          * good hardware PTE piece. Alternatives seem worse.
1186                          */
1187                         unsigned int x; /* Index of HW PMD in soft cluster */
1188                         x = (start >> PMD_SHIFT) & 15;
1189                         *(unsigned long *)__nocache_fix(&pmdp->pmdv[x]) = prompte;
1190                         start += SRMMU_REAL_PMD_SIZE;
1191                         continue;
1192                 }
1193                 ptep = srmmu_pte_offset(__nocache_fix(pmdp), start);
1194                 *(pte_t *)__nocache_fix(ptep) = __pte(prompte);
1195                 start += PAGE_SIZE;
1196         }
1197 }
1198
1199 #define KERNEL_PTE(page_shifted) ((page_shifted)|SRMMU_CACHE|SRMMU_PRIV|SRMMU_VALID)
1200
1201 /* Create a third-level SRMMU 16MB page mapping. */
1202 static void __init do_large_mapping(unsigned long vaddr, unsigned long phys_base)
1203 {
1204         pgd_t *pgdp = pgd_offset_k(vaddr);
1205         unsigned long big_pte;
1206
1207         big_pte = KERNEL_PTE(phys_base >> 4);
1208         *(pgd_t *)__nocache_fix(pgdp) = __pgd(big_pte);
1209 }
1210
1211 /* Map sp_bank entry SP_ENTRY, starting at virtual address VBASE. */
1212 static unsigned long __init map_spbank(unsigned long vbase, int sp_entry)
1213 {
1214         unsigned long pstart = (sp_banks[sp_entry].base_addr & SRMMU_PGDIR_MASK);
1215         unsigned long vstart = (vbase & SRMMU_PGDIR_MASK);
1216         unsigned long vend = SRMMU_PGDIR_ALIGN(vbase + sp_banks[sp_entry].num_bytes);
1217         /* Map "low" memory only */
1218         const unsigned long min_vaddr = PAGE_OFFSET;
1219         const unsigned long max_vaddr = PAGE_OFFSET + SRMMU_MAXMEM;
1220
1221         if (vstart < min_vaddr || vstart >= max_vaddr)
1222                 return vstart;
1223
1224         if (vend > max_vaddr || vend < min_vaddr)
1225                 vend = max_vaddr;
1226
1227         while(vstart < vend) {
1228                 do_large_mapping(vstart, pstart);
1229                 vstart += SRMMU_PGDIR_SIZE; pstart += SRMMU_PGDIR_SIZE;
1230         }
1231         return vstart;
1232 }
1233
1234 static inline void memprobe_error(char *msg)
1235 {
1236         prom_printf(msg);
1237         prom_printf("Halting now...\n");
1238         prom_halt();
1239 }
1240
1241 static inline void map_kernel(void)
1242 {
1243         int i;
1244
1245         if (phys_base > 0) {
1246                 do_large_mapping(PAGE_OFFSET, phys_base);
1247         }
1248
1249         for (i = 0; sp_banks[i].num_bytes != 0; i++) {
1250                 map_spbank((unsigned long)__va(sp_banks[i].base_addr), i);
1251         }
1252
1253         BTFIXUPSET_SIMM13(user_ptrs_per_pgd, PAGE_OFFSET / SRMMU_PGDIR_SIZE);
1254 }
1255
1256 /* Paging initialization on the Sparc Reference MMU. */
1257 extern void sparc_context_init(int);
1258
1259 void (*poke_srmmu)(void) __cpuinitdata = NULL;
1260
1261 extern unsigned long bootmem_init(unsigned long *pages_avail);
1262
1263 void __init srmmu_paging_init(void)
1264 {
1265         int i, cpunode;
1266         char node_str[128];
1267         pgd_t *pgd;
1268         pmd_t *pmd;
1269         pte_t *pte;
1270         unsigned long pages_avail;
1271
1272         sparc_iomap.start = SUN4M_IOBASE_VADDR; /* 16MB of IOSPACE on all sun4m's. */
1273
1274         if (sparc_cpu_model == sun4d)
1275                 num_contexts = 65536; /* We know it is Viking */
1276         else {
1277                 /* Find the number of contexts on the srmmu. */
1278                 cpunode = prom_getchild(prom_root_node);
1279                 num_contexts = 0;
1280                 while(cpunode != 0) {
1281                         prom_getstring(cpunode, "device_type", node_str, sizeof(node_str));
1282                         if(!strcmp(node_str, "cpu")) {
1283                                 num_contexts = prom_getintdefault(cpunode, "mmu-nctx", 0x8);
1284                                 break;
1285                         }
1286                         cpunode = prom_getsibling(cpunode);
1287                 }
1288         }
1289
1290         if(!num_contexts) {
1291                 prom_printf("Something wrong, can't find cpu node in paging_init.\n");
1292                 prom_halt();
1293         }
1294
1295         pages_avail = 0;
1296         last_valid_pfn = bootmem_init(&pages_avail);
1297
1298         srmmu_nocache_calcsize();
1299         srmmu_nocache_init();
1300         srmmu_inherit_prom_mappings(0xfe400000,(LINUX_OPPROM_ENDVM-PAGE_SIZE));
1301         map_kernel();
1302
1303         /* ctx table has to be physically aligned to its size */
1304         srmmu_context_table = (ctxd_t *)__srmmu_get_nocache(num_contexts*sizeof(ctxd_t), num_contexts*sizeof(ctxd_t));
1305         srmmu_ctx_table_phys = (ctxd_t *)__nocache_pa((unsigned long)srmmu_context_table);
1306
1307         for(i = 0; i < num_contexts; i++)
1308                 srmmu_ctxd_set((ctxd_t *)__nocache_fix(&srmmu_context_table[i]), srmmu_swapper_pg_dir);
1309
1310         flush_cache_all();
1311         srmmu_set_ctable_ptr((unsigned long)srmmu_ctx_table_phys);
1312 #ifdef CONFIG_SMP
1313         /* Stop from hanging here... */
1314         local_flush_tlb_all();
1315 #else
1316         flush_tlb_all();
1317 #endif
1318         poke_srmmu();
1319
1320         srmmu_allocate_ptable_skeleton(sparc_iomap.start, IOBASE_END);
1321         srmmu_allocate_ptable_skeleton(DVMA_VADDR, DVMA_END);
1322
1323         srmmu_allocate_ptable_skeleton(
1324                 __fix_to_virt(__end_of_fixed_addresses - 1), FIXADDR_TOP);
1325         srmmu_allocate_ptable_skeleton(PKMAP_BASE, PKMAP_END);
1326
1327         pgd = pgd_offset_k(PKMAP_BASE);
1328         pmd = srmmu_pmd_offset(pgd, PKMAP_BASE);
1329         pte = srmmu_pte_offset(pmd, PKMAP_BASE);
1330         pkmap_page_table = pte;
1331
1332         flush_cache_all();
1333         flush_tlb_all();
1334
1335         sparc_context_init(num_contexts);
1336
1337         kmap_init();
1338
1339         {
1340                 unsigned long zones_size[MAX_NR_ZONES];
1341                 unsigned long zholes_size[MAX_NR_ZONES];
1342                 unsigned long npages;
1343                 int znum;
1344
1345                 for (znum = 0; znum < MAX_NR_ZONES; znum++)
1346                         zones_size[znum] = zholes_size[znum] = 0;
1347
1348                 npages = max_low_pfn - pfn_base;
1349
1350                 zones_size[ZONE_DMA] = npages;
1351                 zholes_size[ZONE_DMA] = npages - pages_avail;
1352
1353                 npages = highend_pfn - max_low_pfn;
1354                 zones_size[ZONE_HIGHMEM] = npages;
1355                 zholes_size[ZONE_HIGHMEM] = npages - calc_highpages();
1356
1357                 free_area_init_node(0, zones_size, pfn_base, zholes_size);
1358         }
1359 }
1360
1361 static void srmmu_mmu_info(struct seq_file *m)
1362 {
1363         seq_printf(m,
1364                    "MMU type\t: %s\n"
1365                    "contexts\t: %d\n"
1366                    "nocache total\t: %ld\n"
1367                    "nocache used\t: %d\n",
1368                    srmmu_name,
1369                    num_contexts,
1370                    srmmu_nocache_size,
1371                    srmmu_nocache_map.used << SRMMU_NOCACHE_BITMAP_SHIFT);
1372 }
1373
1374 static void srmmu_update_mmu_cache(struct vm_area_struct * vma, unsigned long address, pte_t pte)
1375 {
1376 }
1377
1378 static void srmmu_destroy_context(struct mm_struct *mm)
1379 {
1380
1381         if(mm->context != NO_CONTEXT) {
1382                 flush_cache_mm(mm);
1383                 srmmu_ctxd_set(&srmmu_context_table[mm->context], srmmu_swapper_pg_dir);
1384                 flush_tlb_mm(mm);
1385                 spin_lock(&srmmu_context_spinlock);
1386                 free_context(mm->context);
1387                 spin_unlock(&srmmu_context_spinlock);
1388                 mm->context = NO_CONTEXT;
1389         }
1390 }
1391
1392 /* Init various srmmu chip types. */
1393 static void __init srmmu_is_bad(void)
1394 {
1395         prom_printf("Could not determine SRMMU chip type.\n");
1396         prom_halt();
1397 }
1398
1399 static void __init init_vac_layout(void)
1400 {
1401         int nd, cache_lines;
1402         char node_str[128];
1403 #ifdef CONFIG_SMP
1404         int cpu = 0;
1405         unsigned long max_size = 0;
1406         unsigned long min_line_size = 0x10000000;
1407 #endif
1408
1409         nd = prom_getchild(prom_root_node);
1410         while((nd = prom_getsibling(nd)) != 0) {
1411                 prom_getstring(nd, "device_type", node_str, sizeof(node_str));
1412                 if(!strcmp(node_str, "cpu")) {
1413                         vac_line_size = prom_getint(nd, "cache-line-size");
1414                         if (vac_line_size == -1) {
1415                                 prom_printf("can't determine cache-line-size, "
1416                                             "halting.\n");
1417                                 prom_halt();
1418                         }
1419                         cache_lines = prom_getint(nd, "cache-nlines");
1420                         if (cache_lines == -1) {
1421                                 prom_printf("can't determine cache-nlines, halting.\n");
1422                                 prom_halt();
1423                         }
1424
1425                         vac_cache_size = cache_lines * vac_line_size;
1426 #ifdef CONFIG_SMP
1427                         if(vac_cache_size > max_size)
1428                                 max_size = vac_cache_size;
1429                         if(vac_line_size < min_line_size)
1430                                 min_line_size = vac_line_size;
1431                         //FIXME: cpus not contiguous!!
1432                         cpu++;
1433                         if (cpu >= nr_cpu_ids || !cpu_online(cpu))
1434                                 break;
1435 #else
1436                         break;
1437 #endif
1438                 }
1439         }
1440         if(nd == 0) {
1441                 prom_printf("No CPU nodes found, halting.\n");
1442                 prom_halt();
1443         }
1444 #ifdef CONFIG_SMP
1445         vac_cache_size = max_size;
1446         vac_line_size = min_line_size;
1447 #endif
1448         printk("SRMMU: Using VAC size of %d bytes, line size %d bytes.\n",
1449                (int)vac_cache_size, (int)vac_line_size);
1450 }
1451
1452 static void __cpuinit poke_hypersparc(void)
1453 {
1454         volatile unsigned long clear;
1455         unsigned long mreg = srmmu_get_mmureg();
1456
1457         hyper_flush_unconditional_combined();
1458
1459         mreg &= ~(HYPERSPARC_CWENABLE);
1460         mreg |= (HYPERSPARC_CENABLE | HYPERSPARC_WBENABLE);
1461         mreg |= (HYPERSPARC_CMODE);
1462
1463         srmmu_set_mmureg(mreg);
1464
1465 #if 0 /* XXX I think this is bad news... -DaveM */
1466         hyper_clear_all_tags();
1467 #endif
1468
1469         put_ross_icr(HYPERSPARC_ICCR_FTD | HYPERSPARC_ICCR_ICE);
1470         hyper_flush_whole_icache();
1471         clear = srmmu_get_faddr();
1472         clear = srmmu_get_fstatus();
1473 }
1474
1475 static void __init init_hypersparc(void)
1476 {
1477         srmmu_name = "ROSS HyperSparc";
1478         srmmu_modtype = HyperSparc;
1479
1480         init_vac_layout();
1481
1482         is_hypersparc = 1;
1483
1484         BTFIXUPSET_CALL(pte_clear, srmmu_pte_clear, BTFIXUPCALL_NORM);
1485         BTFIXUPSET_CALL(pmd_clear, srmmu_pmd_clear, BTFIXUPCALL_NORM);
1486         BTFIXUPSET_CALL(pgd_clear, srmmu_pgd_clear, BTFIXUPCALL_NORM);
1487         BTFIXUPSET_CALL(flush_cache_all, hypersparc_flush_cache_all, BTFIXUPCALL_NORM);
1488         BTFIXUPSET_CALL(flush_cache_mm, hypersparc_flush_cache_mm, BTFIXUPCALL_NORM);
1489         BTFIXUPSET_CALL(flush_cache_range, hypersparc_flush_cache_range, BTFIXUPCALL_NORM);
1490         BTFIXUPSET_CALL(flush_cache_page, hypersparc_flush_cache_page, BTFIXUPCALL_NORM);
1491
1492         BTFIXUPSET_CALL(flush_tlb_all, hypersparc_flush_tlb_all, BTFIXUPCALL_NORM);
1493         BTFIXUPSET_CALL(flush_tlb_mm, hypersparc_flush_tlb_mm, BTFIXUPCALL_NORM);
1494         BTFIXUPSET_CALL(flush_tlb_range, hypersparc_flush_tlb_range, BTFIXUPCALL_NORM);
1495         BTFIXUPSET_CALL(flush_tlb_page, hypersparc_flush_tlb_page, BTFIXUPCALL_NORM);
1496
1497         BTFIXUPSET_CALL(__flush_page_to_ram, hypersparc_flush_page_to_ram, BTFIXUPCALL_NORM);
1498         BTFIXUPSET_CALL(flush_sig_insns, hypersparc_flush_sig_insns, BTFIXUPCALL_NORM);
1499         BTFIXUPSET_CALL(flush_page_for_dma, hypersparc_flush_page_for_dma, BTFIXUPCALL_NOP);
1500
1501
1502         poke_srmmu = poke_hypersparc;
1503
1504         hypersparc_setup_blockops();
1505 }
1506
1507 static void __cpuinit poke_cypress(void)
1508 {
1509         unsigned long mreg = srmmu_get_mmureg();
1510         unsigned long faddr, tagval;
1511         volatile unsigned long cypress_sucks;
1512         volatile unsigned long clear;
1513
1514         clear = srmmu_get_faddr();
1515         clear = srmmu_get_fstatus();
1516
1517         if (!(mreg & CYPRESS_CENABLE)) {
1518                 for(faddr = 0x0; faddr < 0x10000; faddr += 20) {
1519                         __asm__ __volatile__("sta %%g0, [%0 + %1] %2\n\t"
1520                                              "sta %%g0, [%0] %2\n\t" : :
1521                                              "r" (faddr), "r" (0x40000),
1522                                              "i" (ASI_M_DATAC_TAG));
1523                 }
1524         } else {
1525                 for(faddr = 0; faddr < 0x10000; faddr += 0x20) {
1526                         __asm__ __volatile__("lda [%1 + %2] %3, %0\n\t" :
1527                                              "=r" (tagval) :
1528                                              "r" (faddr), "r" (0x40000),
1529                                              "i" (ASI_M_DATAC_TAG));
1530
1531                         /* If modified and valid, kick it. */
1532                         if((tagval & 0x60) == 0x60)
1533                                 cypress_sucks = *(unsigned long *)
1534                                                         (0xf0020000 + faddr);
1535                 }
1536         }
1537
1538         /* And one more, for our good neighbor, Mr. Broken Cypress. */
1539         clear = srmmu_get_faddr();
1540         clear = srmmu_get_fstatus();
1541
1542         mreg |= (CYPRESS_CENABLE | CYPRESS_CMODE);
1543         srmmu_set_mmureg(mreg);
1544 }
1545
1546 static void __init init_cypress_common(void)
1547 {
1548         init_vac_layout();
1549
1550         BTFIXUPSET_CALL(pte_clear, srmmu_pte_clear, BTFIXUPCALL_NORM);
1551         BTFIXUPSET_CALL(pmd_clear, srmmu_pmd_clear, BTFIXUPCALL_NORM);
1552         BTFIXUPSET_CALL(pgd_clear, srmmu_pgd_clear, BTFIXUPCALL_NORM);
1553         BTFIXUPSET_CALL(flush_cache_all, cypress_flush_cache_all, BTFIXUPCALL_NORM);
1554         BTFIXUPSET_CALL(flush_cache_mm, cypress_flush_cache_mm, BTFIXUPCALL_NORM);
1555         BTFIXUPSET_CALL(flush_cache_range, cypress_flush_cache_range, BTFIXUPCALL_NORM);
1556         BTFIXUPSET_CALL(flush_cache_page, cypress_flush_cache_page, BTFIXUPCALL_NORM);
1557
1558         BTFIXUPSET_CALL(flush_tlb_all, cypress_flush_tlb_all, BTFIXUPCALL_NORM);
1559         BTFIXUPSET_CALL(flush_tlb_mm, cypress_flush_tlb_mm, BTFIXUPCALL_NORM);
1560         BTFIXUPSET_CALL(flush_tlb_page, cypress_flush_tlb_page, BTFIXUPCALL_NORM);
1561         BTFIXUPSET_CALL(flush_tlb_range, cypress_flush_tlb_range, BTFIXUPCALL_NORM);
1562
1563
1564         BTFIXUPSET_CALL(__flush_page_to_ram, cypress_flush_page_to_ram, BTFIXUPCALL_NORM);
1565         BTFIXUPSET_CALL(flush_sig_insns, cypress_flush_sig_insns, BTFIXUPCALL_NOP);
1566         BTFIXUPSET_CALL(flush_page_for_dma, cypress_flush_page_for_dma, BTFIXUPCALL_NOP);
1567
1568         poke_srmmu = poke_cypress;
1569 }
1570
1571 static void __init init_cypress_604(void)
1572 {
1573         srmmu_name = "ROSS Cypress-604(UP)";
1574         srmmu_modtype = Cypress;
1575         init_cypress_common();
1576 }
1577
1578 static void __init init_cypress_605(unsigned long mrev)
1579 {
1580         srmmu_name = "ROSS Cypress-605(MP)";
1581         if(mrev == 0xe) {
1582                 srmmu_modtype = Cypress_vE;
1583                 hwbug_bitmask |= HWBUG_COPYBACK_BROKEN;
1584         } else {
1585                 if(mrev == 0xd) {
1586                         srmmu_modtype = Cypress_vD;
1587                         hwbug_bitmask |= HWBUG_ASIFLUSH_BROKEN;
1588                 } else {
1589                         srmmu_modtype = Cypress;
1590                 }
1591         }
1592         init_cypress_common();
1593 }
1594
1595 static void __cpuinit poke_swift(void)
1596 {
1597         unsigned long mreg;
1598
1599         /* Clear any crap from the cache or else... */
1600         swift_flush_cache_all();
1601
1602         /* Enable I & D caches */
1603         mreg = srmmu_get_mmureg();
1604         mreg |= (SWIFT_IE | SWIFT_DE);
1605         /*
1606          * The Swift branch folding logic is completely broken.  At
1607          * trap time, if things are just right, if can mistakenly
1608          * think that a trap is coming from kernel mode when in fact
1609          * it is coming from user mode (it mis-executes the branch in
1610          * the trap code).  So you see things like crashme completely
1611          * hosing your machine which is completely unacceptable.  Turn
1612          * this shit off... nice job Fujitsu.
1613          */
1614         mreg &= ~(SWIFT_BF);
1615         srmmu_set_mmureg(mreg);
1616 }
1617
1618 #define SWIFT_MASKID_ADDR  0x10003018
1619 static void __init init_swift(void)
1620 {
1621         unsigned long swift_rev;
1622
1623         __asm__ __volatile__("lda [%1] %2, %0\n\t"
1624                              "srl %0, 0x18, %0\n\t" :
1625                              "=r" (swift_rev) :
1626                              "r" (SWIFT_MASKID_ADDR), "i" (ASI_M_BYPASS));
1627         srmmu_name = "Fujitsu Swift";
1628         switch(swift_rev) {
1629         case 0x11:
1630         case 0x20:
1631         case 0x23:
1632         case 0x30:
1633                 srmmu_modtype = Swift_lots_o_bugs;
1634                 hwbug_bitmask |= (HWBUG_KERN_ACCBROKEN | HWBUG_KERN_CBITBROKEN);
1635                 /*
1636                  * Gee george, I wonder why Sun is so hush hush about
1637                  * this hardware bug... really braindamage stuff going
1638                  * on here.  However I think we can find a way to avoid
1639                  * all of the workaround overhead under Linux.  Basically,
1640                  * any page fault can cause kernel pages to become user
1641                  * accessible (the mmu gets confused and clears some of
1642                  * the ACC bits in kernel ptes).  Aha, sounds pretty
1643                  * horrible eh?  But wait, after extensive testing it appears
1644                  * that if you use pgd_t level large kernel pte's (like the
1645                  * 4MB pages on the Pentium) the bug does not get tripped
1646                  * at all.  This avoids almost all of the major overhead.
1647                  * Welcome to a world where your vendor tells you to,
1648                  * "apply this kernel patch" instead of "sorry for the
1649                  * broken hardware, send it back and we'll give you
1650                  * properly functioning parts"
1651                  */
1652                 break;
1653         case 0x25:
1654         case 0x31:
1655                 srmmu_modtype = Swift_bad_c;
1656                 hwbug_bitmask |= HWBUG_KERN_CBITBROKEN;
1657                 /*
1658                  * You see Sun allude to this hardware bug but never
1659                  * admit things directly, they'll say things like,
1660                  * "the Swift chip cache problems" or similar.
1661                  */
1662                 break;
1663         default:
1664                 srmmu_modtype = Swift_ok;
1665                 break;
1666         };
1667
1668         BTFIXUPSET_CALL(flush_cache_all, swift_flush_cache_all, BTFIXUPCALL_NORM);
1669         BTFIXUPSET_CALL(flush_cache_mm, swift_flush_cache_mm, BTFIXUPCALL_NORM);
1670         BTFIXUPSET_CALL(flush_cache_page, swift_flush_cache_page, BTFIXUPCALL_NORM);
1671         BTFIXUPSET_CALL(flush_cache_range, swift_flush_cache_range, BTFIXUPCALL_NORM);
1672
1673
1674         BTFIXUPSET_CALL(flush_tlb_all, swift_flush_tlb_all, BTFIXUPCALL_NORM);
1675         BTFIXUPSET_CALL(flush_tlb_mm, swift_flush_tlb_mm, BTFIXUPCALL_NORM);
1676         BTFIXUPSET_CALL(flush_tlb_page, swift_flush_tlb_page, BTFIXUPCALL_NORM);
1677         BTFIXUPSET_CALL(flush_tlb_range, swift_flush_tlb_range, BTFIXUPCALL_NORM);
1678
1679         BTFIXUPSET_CALL(__flush_page_to_ram, swift_flush_page_to_ram, BTFIXUPCALL_NORM);
1680         BTFIXUPSET_CALL(flush_sig_insns, swift_flush_sig_insns, BTFIXUPCALL_NORM);
1681         BTFIXUPSET_CALL(flush_page_for_dma, swift_flush_page_for_dma, BTFIXUPCALL_NORM);
1682
1683         BTFIXUPSET_CALL(update_mmu_cache, swift_update_mmu_cache, BTFIXUPCALL_NORM);
1684
1685         flush_page_for_dma_global = 0;
1686
1687         /*
1688          * Are you now convinced that the Swift is one of the
1689          * biggest VLSI abortions of all time?  Bravo Fujitsu!
1690          * Fujitsu, the !#?!%$'d up processor people.  I bet if
1691          * you examined the microcode of the Swift you'd find
1692          * XXX's all over the place.
1693          */
1694         poke_srmmu = poke_swift;
1695 }
1696
1697 static void turbosparc_flush_cache_all(void)
1698 {
1699         flush_user_windows();
1700         turbosparc_idflash_clear();
1701 }
1702
1703 static void turbosparc_flush_cache_mm(struct mm_struct *mm)
1704 {
1705         FLUSH_BEGIN(mm)
1706         flush_user_windows();
1707         turbosparc_idflash_clear();
1708         FLUSH_END
1709 }
1710
1711 static void turbosparc_flush_cache_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
1712 {
1713         FLUSH_BEGIN(vma->vm_mm)
1714         flush_user_windows();
1715         turbosparc_idflash_clear();
1716         FLUSH_END
1717 }
1718
1719 static void turbosparc_flush_cache_page(struct vm_area_struct *vma, unsigned long page)
1720 {
1721         FLUSH_BEGIN(vma->vm_mm)
1722         flush_user_windows();
1723         if (vma->vm_flags & VM_EXEC)
1724                 turbosparc_flush_icache();
1725         turbosparc_flush_dcache();
1726         FLUSH_END
1727 }
1728
1729 /* TurboSparc is copy-back, if we turn it on, but this does not work. */
1730 static void turbosparc_flush_page_to_ram(unsigned long page)
1731 {
1732 #ifdef TURBOSPARC_WRITEBACK
1733         volatile unsigned long clear;
1734
1735         if (srmmu_hwprobe(page))
1736                 turbosparc_flush_page_cache(page);
1737         clear = srmmu_get_fstatus();
1738 #endif
1739 }
1740
1741 static void turbosparc_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr)
1742 {
1743 }
1744
1745 static void turbosparc_flush_page_for_dma(unsigned long page)
1746 {
1747         turbosparc_flush_dcache();
1748 }
1749
1750 static void turbosparc_flush_tlb_all(void)
1751 {
1752         srmmu_flush_whole_tlb();
1753 }
1754
1755 static void turbosparc_flush_tlb_mm(struct mm_struct *mm)
1756 {
1757         FLUSH_BEGIN(mm)
1758         srmmu_flush_whole_tlb();
1759         FLUSH_END
1760 }
1761
1762 static void turbosparc_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
1763 {
1764         FLUSH_BEGIN(vma->vm_mm)
1765         srmmu_flush_whole_tlb();
1766         FLUSH_END
1767 }
1768
1769 static void turbosparc_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
1770 {
1771         FLUSH_BEGIN(vma->vm_mm)
1772         srmmu_flush_whole_tlb();
1773         FLUSH_END
1774 }
1775
1776
1777 static void __cpuinit poke_turbosparc(void)
1778 {
1779         unsigned long mreg = srmmu_get_mmureg();
1780         unsigned long ccreg;
1781
1782         /* Clear any crap from the cache or else... */
1783         turbosparc_flush_cache_all();
1784         mreg &= ~(TURBOSPARC_ICENABLE | TURBOSPARC_DCENABLE); /* Temporarily disable I & D caches */
1785         mreg &= ~(TURBOSPARC_PCENABLE);         /* Don't check parity */
1786         srmmu_set_mmureg(mreg);
1787
1788         ccreg = turbosparc_get_ccreg();
1789
1790 #ifdef TURBOSPARC_WRITEBACK
1791         ccreg |= (TURBOSPARC_SNENABLE);         /* Do DVMA snooping in Dcache */
1792         ccreg &= ~(TURBOSPARC_uS2 | TURBOSPARC_WTENABLE);
1793                         /* Write-back D-cache, emulate VLSI
1794                          * abortion number three, not number one */
1795 #else
1796         /* For now let's play safe, optimize later */
1797         ccreg |= (TURBOSPARC_SNENABLE | TURBOSPARC_WTENABLE);
1798                         /* Do DVMA snooping in Dcache, Write-thru D-cache */
1799         ccreg &= ~(TURBOSPARC_uS2);
1800                         /* Emulate VLSI abortion number three, not number one */
1801 #endif
1802
1803         switch (ccreg & 7) {
1804         case 0: /* No SE cache */
1805         case 7: /* Test mode */
1806                 break;
1807         default:
1808                 ccreg |= (TURBOSPARC_SCENABLE);
1809         }
1810         turbosparc_set_ccreg (ccreg);
1811
1812         mreg |= (TURBOSPARC_ICENABLE | TURBOSPARC_DCENABLE); /* I & D caches on */
1813         mreg |= (TURBOSPARC_ICSNOOP);           /* Icache snooping on */
1814         srmmu_set_mmureg(mreg);
1815 }
1816
1817 static void __init init_turbosparc(void)
1818 {
1819         srmmu_name = "Fujitsu TurboSparc";
1820         srmmu_modtype = TurboSparc;
1821
1822         BTFIXUPSET_CALL(flush_cache_all, turbosparc_flush_cache_all, BTFIXUPCALL_NORM);
1823         BTFIXUPSET_CALL(flush_cache_mm, turbosparc_flush_cache_mm, BTFIXUPCALL_NORM);
1824         BTFIXUPSET_CALL(flush_cache_page, turbosparc_flush_cache_page, BTFIXUPCALL_NORM);
1825         BTFIXUPSET_CALL(flush_cache_range, turbosparc_flush_cache_range, BTFIXUPCALL_NORM);
1826
1827         BTFIXUPSET_CALL(flush_tlb_all, turbosparc_flush_tlb_all, BTFIXUPCALL_NORM);
1828         BTFIXUPSET_CALL(flush_tlb_mm, turbosparc_flush_tlb_mm, BTFIXUPCALL_NORM);
1829         BTFIXUPSET_CALL(flush_tlb_page, turbosparc_flush_tlb_page, BTFIXUPCALL_NORM);
1830         BTFIXUPSET_CALL(flush_tlb_range, turbosparc_flush_tlb_range, BTFIXUPCALL_NORM);
1831
1832         BTFIXUPSET_CALL(__flush_page_to_ram, turbosparc_flush_page_to_ram, BTFIXUPCALL_NORM);
1833
1834         BTFIXUPSET_CALL(flush_sig_insns, turbosparc_flush_sig_insns, BTFIXUPCALL_NOP);
1835         BTFIXUPSET_CALL(flush_page_for_dma, turbosparc_flush_page_for_dma, BTFIXUPCALL_NORM);
1836
1837         poke_srmmu = poke_turbosparc;
1838 }
1839
1840 static void __cpuinit poke_tsunami(void)
1841 {
1842         unsigned long mreg = srmmu_get_mmureg();
1843
1844         tsunami_flush_icache();
1845         tsunami_flush_dcache();
1846         mreg &= ~TSUNAMI_ITD;
1847         mreg |= (TSUNAMI_IENAB | TSUNAMI_DENAB);
1848         srmmu_set_mmureg(mreg);
1849 }
1850
1851 static void __init init_tsunami(void)
1852 {
1853         /*
1854          * Tsunami's pretty sane, Sun and TI actually got it
1855          * somewhat right this time.  Fujitsu should have
1856          * taken some lessons from them.
1857          */
1858
1859         srmmu_name = "TI Tsunami";
1860         srmmu_modtype = Tsunami;
1861
1862         BTFIXUPSET_CALL(flush_cache_all, tsunami_flush_cache_all, BTFIXUPCALL_NORM);
1863         BTFIXUPSET_CALL(flush_cache_mm, tsunami_flush_cache_mm, BTFIXUPCALL_NORM);
1864         BTFIXUPSET_CALL(flush_cache_page, tsunami_flush_cache_page, BTFIXUPCALL_NORM);
1865         BTFIXUPSET_CALL(flush_cache_range, tsunami_flush_cache_range, BTFIXUPCALL_NORM);
1866
1867
1868         BTFIXUPSET_CALL(flush_tlb_all, tsunami_flush_tlb_all, BTFIXUPCALL_NORM);
1869         BTFIXUPSET_CALL(flush_tlb_mm, tsunami_flush_tlb_mm, BTFIXUPCALL_NORM);
1870         BTFIXUPSET_CALL(flush_tlb_page, tsunami_flush_tlb_page, BTFIXUPCALL_NORM);
1871         BTFIXUPSET_CALL(flush_tlb_range, tsunami_flush_tlb_range, BTFIXUPCALL_NORM);
1872
1873         BTFIXUPSET_CALL(__flush_page_to_ram, tsunami_flush_page_to_ram, BTFIXUPCALL_NOP);
1874         BTFIXUPSET_CALL(flush_sig_insns, tsunami_flush_sig_insns, BTFIXUPCALL_NORM);
1875         BTFIXUPSET_CALL(flush_page_for_dma, tsunami_flush_page_for_dma, BTFIXUPCALL_NORM);
1876
1877         poke_srmmu = poke_tsunami;
1878
1879         tsunami_setup_blockops();
1880 }
1881
1882 static void __cpuinit poke_viking(void)
1883 {
1884         unsigned long mreg = srmmu_get_mmureg();
1885         static int smp_catch;
1886
1887         if(viking_mxcc_present) {
1888                 unsigned long mxcc_control = mxcc_get_creg();
1889
1890                 mxcc_control |= (MXCC_CTL_ECE | MXCC_CTL_PRE | MXCC_CTL_MCE);
1891                 mxcc_control &= ~(MXCC_CTL_RRC);
1892                 mxcc_set_creg(mxcc_control);
1893
1894                 /*
1895                  * We don't need memory parity checks.
1896                  * XXX This is a mess, have to dig out later. ecd.
1897                 viking_mxcc_turn_off_parity(&mreg, &mxcc_control);
1898                  */
1899
1900                 /* We do cache ptables on MXCC. */
1901                 mreg |= VIKING_TCENABLE;
1902         } else {
1903                 unsigned long bpreg;
1904
1905                 mreg &= ~(VIKING_TCENABLE);
1906                 if(smp_catch++) {
1907                         /* Must disable mixed-cmd mode here for other cpu's. */
1908                         bpreg = viking_get_bpreg();
1909                         bpreg &= ~(VIKING_ACTION_MIX);
1910                         viking_set_bpreg(bpreg);
1911
1912                         /* Just in case PROM does something funny. */
1913                         msi_set_sync();
1914                 }
1915         }
1916
1917         mreg |= VIKING_SPENABLE;
1918         mreg |= (VIKING_ICENABLE | VIKING_DCENABLE);
1919         mreg |= VIKING_SBENABLE;
1920         mreg &= ~(VIKING_ACENABLE);
1921         srmmu_set_mmureg(mreg);
1922 }
1923
1924 static void __init init_viking(void)
1925 {
1926         unsigned long mreg = srmmu_get_mmureg();
1927
1928         /* Ahhh, the viking.  SRMMU VLSI abortion number two... */
1929         if(mreg & VIKING_MMODE) {
1930                 srmmu_name = "TI Viking";
1931                 viking_mxcc_present = 0;
1932                 msi_set_sync();
1933
1934                 BTFIXUPSET_CALL(pte_clear, srmmu_pte_clear, BTFIXUPCALL_NORM);
1935                 BTFIXUPSET_CALL(pmd_clear, srmmu_pmd_clear, BTFIXUPCALL_NORM);
1936                 BTFIXUPSET_CALL(pgd_clear, srmmu_pgd_clear, BTFIXUPCALL_NORM);
1937
1938                 /*
1939                  * We need this to make sure old viking takes no hits
1940                  * on it's cache for dma snoops to workaround the
1941                  * "load from non-cacheable memory" interrupt bug.
1942                  * This is only necessary because of the new way in
1943                  * which we use the IOMMU.
1944                  */
1945                 BTFIXUPSET_CALL(flush_page_for_dma, viking_flush_page, BTFIXUPCALL_NORM);
1946
1947                 flush_page_for_dma_global = 0;
1948         } else {
1949                 srmmu_name = "TI Viking/MXCC";
1950                 viking_mxcc_present = 1;
1951
1952                 srmmu_cache_pagetables = 1;
1953
1954                 /* MXCC vikings lack the DMA snooping bug. */
1955                 BTFIXUPSET_CALL(flush_page_for_dma, viking_flush_page_for_dma, BTFIXUPCALL_NOP);
1956         }
1957
1958         BTFIXUPSET_CALL(flush_cache_all, viking_flush_cache_all, BTFIXUPCALL_NORM);
1959         BTFIXUPSET_CALL(flush_cache_mm, viking_flush_cache_mm, BTFIXUPCALL_NORM);
1960         BTFIXUPSET_CALL(flush_cache_page, viking_flush_cache_page, BTFIXUPCALL_NORM);
1961         BTFIXUPSET_CALL(flush_cache_range, viking_flush_cache_range, BTFIXUPCALL_NORM);
1962
1963 #ifdef CONFIG_SMP
1964         if (sparc_cpu_model == sun4d) {
1965                 BTFIXUPSET_CALL(flush_tlb_all, sun4dsmp_flush_tlb_all, BTFIXUPCALL_NORM);
1966                 BTFIXUPSET_CALL(flush_tlb_mm, sun4dsmp_flush_tlb_mm, BTFIXUPCALL_NORM);
1967                 BTFIXUPSET_CALL(flush_tlb_page, sun4dsmp_flush_tlb_page, BTFIXUPCALL_NORM);
1968                 BTFIXUPSET_CALL(flush_tlb_range, sun4dsmp_flush_tlb_range, BTFIXUPCALL_NORM);
1969         } else
1970 #endif
1971         {
1972                 BTFIXUPSET_CALL(flush_tlb_all, viking_flush_tlb_all, BTFIXUPCALL_NORM);
1973                 BTFIXUPSET_CALL(flush_tlb_mm, viking_flush_tlb_mm, BTFIXUPCALL_NORM);
1974                 BTFIXUPSET_CALL(flush_tlb_page, viking_flush_tlb_page, BTFIXUPCALL_NORM);
1975                 BTFIXUPSET_CALL(flush_tlb_range, viking_flush_tlb_range, BTFIXUPCALL_NORM);
1976         }
1977
1978         BTFIXUPSET_CALL(__flush_page_to_ram, viking_flush_page_to_ram, BTFIXUPCALL_NOP);
1979         BTFIXUPSET_CALL(flush_sig_insns, viking_flush_sig_insns, BTFIXUPCALL_NOP);
1980
1981         poke_srmmu = poke_viking;
1982 }
1983
1984 #ifdef CONFIG_SPARC_LEON
1985
1986 void __init poke_leonsparc(void)
1987 {
1988 }
1989
1990 void __init init_leon(void)
1991 {
1992
1993         srmmu_name = "LEON";
1994
1995         BTFIXUPSET_CALL(flush_cache_all, leon_flush_cache_all,
1996                         BTFIXUPCALL_NORM);
1997         BTFIXUPSET_CALL(flush_cache_mm, leon_flush_cache_all,
1998                         BTFIXUPCALL_NORM);
1999         BTFIXUPSET_CALL(flush_cache_page, leon_flush_pcache_all,
2000                         BTFIXUPCALL_NORM);
2001         BTFIXUPSET_CALL(flush_cache_range, leon_flush_cache_all,
2002                         BTFIXUPCALL_NORM);
2003         BTFIXUPSET_CALL(flush_page_for_dma, leon_flush_dcache_all,
2004                         BTFIXUPCALL_NORM);
2005
2006         BTFIXUPSET_CALL(flush_tlb_all, leon_flush_tlb_all, BTFIXUPCALL_NORM);
2007         BTFIXUPSET_CALL(flush_tlb_mm, leon_flush_tlb_all, BTFIXUPCALL_NORM);
2008         BTFIXUPSET_CALL(flush_tlb_page, leon_flush_tlb_all, BTFIXUPCALL_NORM);
2009         BTFIXUPSET_CALL(flush_tlb_range, leon_flush_tlb_all, BTFIXUPCALL_NORM);
2010
2011         BTFIXUPSET_CALL(__flush_page_to_ram, leon_flush_cache_all,
2012                         BTFIXUPCALL_NOP);
2013         BTFIXUPSET_CALL(flush_sig_insns, leon_flush_cache_all, BTFIXUPCALL_NOP);
2014
2015         poke_srmmu = poke_leonsparc;
2016
2017         srmmu_cache_pagetables = 0;
2018
2019         leon_flush_during_switch = leon_flush_needed();
2020 }
2021 #endif
2022
2023 /* Probe for the srmmu chip version. */
2024 static void __init get_srmmu_type(void)
2025 {
2026         unsigned long mreg, psr;
2027         unsigned long mod_typ, mod_rev, psr_typ, psr_vers;
2028
2029         srmmu_modtype = SRMMU_INVAL_MOD;
2030         hwbug_bitmask = 0;
2031
2032         mreg = srmmu_get_mmureg(); psr = get_psr();
2033         mod_typ = (mreg & 0xf0000000) >> 28;
2034         mod_rev = (mreg & 0x0f000000) >> 24;
2035         psr_typ = (psr >> 28) & 0xf;
2036         psr_vers = (psr >> 24) & 0xf;
2037
2038         /* First, check for sparc-leon. */
2039         if (sparc_cpu_model == sparc_leon) {
2040                 init_leon();
2041                 return;
2042         }
2043
2044         /* Second, check for HyperSparc or Cypress. */
2045         if(mod_typ == 1) {
2046                 switch(mod_rev) {
2047                 case 7:
2048                         /* UP or MP Hypersparc */
2049                         init_hypersparc();
2050                         break;
2051                 case 0:
2052                 case 2:
2053                         /* Uniprocessor Cypress */
2054                         init_cypress_604();
2055                         break;
2056                 case 10:
2057                 case 11:
2058                 case 12:
2059                         /* _REALLY OLD_ Cypress MP chips... */
2060                 case 13:
2061                 case 14:
2062                 case 15:
2063                         /* MP Cypress mmu/cache-controller */
2064                         init_cypress_605(mod_rev);
2065                         break;
2066                 default:
2067                         /* Some other Cypress revision, assume a 605. */
2068                         init_cypress_605(mod_rev);
2069                         break;
2070                 };
2071                 return;
2072         }
2073
2074         /*
2075          * Now Fujitsu TurboSparc. It might happen that it is
2076          * in Swift emulation mode, so we will check later...
2077          */
2078         if (psr_typ == 0 && psr_vers == 5) {
2079                 init_turbosparc();
2080                 return;
2081         }
2082
2083         /* Next check for Fujitsu Swift. */
2084         if(psr_typ == 0 && psr_vers == 4) {
2085                 int cpunode;
2086                 char node_str[128];
2087
2088                 /* Look if it is not a TurboSparc emulating Swift... */
2089                 cpunode = prom_getchild(prom_root_node);
2090                 while((cpunode = prom_getsibling(cpunode)) != 0) {
2091                         prom_getstring(cpunode, "device_type", node_str, sizeof(node_str));
2092                         if(!strcmp(node_str, "cpu")) {
2093                                 if (!prom_getintdefault(cpunode, "psr-implementation", 1) &&
2094                                     prom_getintdefault(cpunode, "psr-version", 1) == 5) {
2095                                         init_turbosparc();
2096                                         return;
2097                                 }
2098                                 break;
2099                         }
2100                 }
2101
2102                 init_swift();
2103                 return;
2104         }
2105
2106         /* Now the Viking family of srmmu. */
2107         if(psr_typ == 4 &&
2108            ((psr_vers == 0) ||
2109             ((psr_vers == 1) && (mod_typ == 0) && (mod_rev == 0)))) {
2110                 init_viking();
2111                 return;
2112         }
2113
2114         /* Finally the Tsunami. */
2115         if(psr_typ == 4 && psr_vers == 1 && (mod_typ || mod_rev)) {
2116                 init_tsunami();
2117                 return;
2118         }
2119
2120         /* Oh well */
2121         srmmu_is_bad();
2122 }
2123
2124 /* don't laugh, static pagetables */
2125 static void srmmu_check_pgt_cache(int low, int high)
2126 {
2127 }
2128
2129 extern unsigned long spwin_mmu_patchme, fwin_mmu_patchme,
2130         tsetup_mmu_patchme, rtrap_mmu_patchme;
2131
2132 extern unsigned long spwin_srmmu_stackchk, srmmu_fwin_stackchk,
2133         tsetup_srmmu_stackchk, srmmu_rett_stackchk;
2134
2135 extern unsigned long srmmu_fault;
2136
2137 #define PATCH_BRANCH(insn, dest) do { \
2138                 iaddr = &(insn); \
2139                 daddr = &(dest); \
2140                 *iaddr = SPARC_BRANCH((unsigned long) daddr, (unsigned long) iaddr); \
2141         } while(0)
2142
2143 static void __init patch_window_trap_handlers(void)
2144 {
2145         unsigned long *iaddr, *daddr;
2146
2147         PATCH_BRANCH(spwin_mmu_patchme, spwin_srmmu_stackchk);
2148         PATCH_BRANCH(fwin_mmu_patchme, srmmu_fwin_stackchk);
2149         PATCH_BRANCH(tsetup_mmu_patchme, tsetup_srmmu_stackchk);
2150         PATCH_BRANCH(rtrap_mmu_patchme, srmmu_rett_stackchk);
2151         PATCH_BRANCH(sparc_ttable[SP_TRAP_TFLT].inst_three, srmmu_fault);
2152         PATCH_BRANCH(sparc_ttable[SP_TRAP_DFLT].inst_three, srmmu_fault);
2153         PATCH_BRANCH(sparc_ttable[SP_TRAP_DACC].inst_three, srmmu_fault);
2154 }
2155
2156 #ifdef CONFIG_SMP
2157 /* Local cross-calls. */
2158 static void smp_flush_page_for_dma(unsigned long page)
2159 {
2160         xc1((smpfunc_t) BTFIXUP_CALL(local_flush_page_for_dma), page);
2161         local_flush_page_for_dma(page);
2162 }
2163
2164 #endif
2165
2166 static pte_t srmmu_pgoff_to_pte(unsigned long pgoff)
2167 {
2168         return __pte((pgoff << SRMMU_PTE_FILE_SHIFT) | SRMMU_FILE);
2169 }
2170
2171 static unsigned long srmmu_pte_to_pgoff(pte_t pte)
2172 {
2173         return pte_val(pte) >> SRMMU_PTE_FILE_SHIFT;
2174 }
2175
2176 static pgprot_t srmmu_pgprot_noncached(pgprot_t prot)
2177 {
2178         prot &= ~__pgprot(SRMMU_CACHE);
2179
2180         return prot;
2181 }
2182
2183 /* Load up routines and constants for sun4m and sun4d mmu */
2184 void __init ld_mmu_srmmu(void)
2185 {
2186         extern void ld_mmu_iommu(void);
2187         extern void ld_mmu_iounit(void);
2188         extern void ___xchg32_sun4md(void);
2189
2190         BTFIXUPSET_SIMM13(pgdir_shift, SRMMU_PGDIR_SHIFT);
2191         BTFIXUPSET_SETHI(pgdir_size, SRMMU_PGDIR_SIZE);
2192         BTFIXUPSET_SETHI(pgdir_mask, SRMMU_PGDIR_MASK);
2193
2194         BTFIXUPSET_SIMM13(ptrs_per_pmd, SRMMU_PTRS_PER_PMD);
2195         BTFIXUPSET_SIMM13(ptrs_per_pgd, SRMMU_PTRS_PER_PGD);
2196
2197         BTFIXUPSET_INT(page_none, pgprot_val(SRMMU_PAGE_NONE));
2198         PAGE_SHARED = pgprot_val(SRMMU_PAGE_SHARED);
2199         BTFIXUPSET_INT(page_copy, pgprot_val(SRMMU_PAGE_COPY));
2200         BTFIXUPSET_INT(page_readonly, pgprot_val(SRMMU_PAGE_RDONLY));
2201         BTFIXUPSET_INT(page_kernel, pgprot_val(SRMMU_PAGE_KERNEL));
2202         page_kernel = pgprot_val(SRMMU_PAGE_KERNEL);
2203
2204         /* Functions */
2205         BTFIXUPSET_CALL(pgprot_noncached, srmmu_pgprot_noncached, BTFIXUPCALL_NORM);
2206 #ifndef CONFIG_SMP
2207         BTFIXUPSET_CALL(___xchg32, ___xchg32_sun4md, BTFIXUPCALL_SWAPG1G2);
2208 #endif
2209         BTFIXUPSET_CALL(do_check_pgt_cache, srmmu_check_pgt_cache, BTFIXUPCALL_NOP);
2210
2211         BTFIXUPSET_CALL(set_pte, srmmu_set_pte, BTFIXUPCALL_SWAPO0O1);
2212         BTFIXUPSET_CALL(switch_mm, srmmu_switch_mm, BTFIXUPCALL_NORM);
2213
2214         BTFIXUPSET_CALL(pte_pfn, srmmu_pte_pfn, BTFIXUPCALL_NORM);
2215         BTFIXUPSET_CALL(pmd_page, srmmu_pmd_page, BTFIXUPCALL_NORM);
2216         BTFIXUPSET_CALL(pgd_page_vaddr, srmmu_pgd_page, BTFIXUPCALL_NORM);
2217
2218         BTFIXUPSET_SETHI(none_mask, 0xF0000000);
2219
2220         BTFIXUPSET_CALL(pte_present, srmmu_pte_present, BTFIXUPCALL_NORM);
2221         BTFIXUPSET_CALL(pte_clear, srmmu_pte_clear, BTFIXUPCALL_SWAPO0G0);
2222
2223         BTFIXUPSET_CALL(pmd_bad, srmmu_pmd_bad, BTFIXUPCALL_NORM);
2224         BTFIXUPSET_CALL(pmd_present, srmmu_pmd_present, BTFIXUPCALL_NORM);
2225         BTFIXUPSET_CALL(pmd_clear, srmmu_pmd_clear, BTFIXUPCALL_SWAPO0G0);
2226
2227         BTFIXUPSET_CALL(pgd_none, srmmu_pgd_none, BTFIXUPCALL_NORM);
2228         BTFIXUPSET_CALL(pgd_bad, srmmu_pgd_bad, BTFIXUPCALL_NORM);
2229         BTFIXUPSET_CALL(pgd_present, srmmu_pgd_present, BTFIXUPCALL_NORM);
2230         BTFIXUPSET_CALL(pgd_clear, srmmu_pgd_clear, BTFIXUPCALL_SWAPO0G0);
2231
2232         BTFIXUPSET_CALL(mk_pte, srmmu_mk_pte, BTFIXUPCALL_NORM);
2233         BTFIXUPSET_CALL(mk_pte_phys, srmmu_mk_pte_phys, BTFIXUPCALL_NORM);
2234         BTFIXUPSET_CALL(mk_pte_io, srmmu_mk_pte_io, BTFIXUPCALL_NORM);
2235         BTFIXUPSET_CALL(pgd_set, srmmu_pgd_set, BTFIXUPCALL_NORM);
2236         BTFIXUPSET_CALL(pmd_set, srmmu_pmd_set, BTFIXUPCALL_NORM);
2237         BTFIXUPSET_CALL(pmd_populate, srmmu_pmd_populate, BTFIXUPCALL_NORM);
2238
2239         BTFIXUPSET_INT(pte_modify_mask, SRMMU_CHG_MASK);
2240         BTFIXUPSET_CALL(pmd_offset, srmmu_pmd_offset, BTFIXUPCALL_NORM);
2241         BTFIXUPSET_CALL(pte_offset_kernel, srmmu_pte_offset, BTFIXUPCALL_NORM);
2242
2243         BTFIXUPSET_CALL(free_pte_fast, srmmu_free_pte_fast, BTFIXUPCALL_NORM);
2244         BTFIXUPSET_CALL(pte_free, srmmu_pte_free, BTFIXUPCALL_NORM);
2245         BTFIXUPSET_CALL(pte_alloc_one_kernel, srmmu_pte_alloc_one_kernel, BTFIXUPCALL_NORM);
2246         BTFIXUPSET_CALL(pte_alloc_one, srmmu_pte_alloc_one, BTFIXUPCALL_NORM);
2247         BTFIXUPSET_CALL(free_pmd_fast, srmmu_pmd_free, BTFIXUPCALL_NORM);
2248         BTFIXUPSET_CALL(pmd_alloc_one, srmmu_pmd_alloc_one, BTFIXUPCALL_NORM);
2249         BTFIXUPSET_CALL(free_pgd_fast, srmmu_free_pgd_fast, BTFIXUPCALL_NORM);
2250         BTFIXUPSET_CALL(get_pgd_fast, srmmu_get_pgd_fast, BTFIXUPCALL_NORM);
2251
2252         BTFIXUPSET_HALF(pte_writei, SRMMU_WRITE);
2253         BTFIXUPSET_HALF(pte_dirtyi, SRMMU_DIRTY);
2254         BTFIXUPSET_HALF(pte_youngi, SRMMU_REF);
2255         BTFIXUPSET_HALF(pte_filei, SRMMU_FILE);
2256         BTFIXUPSET_HALF(pte_wrprotecti, SRMMU_WRITE);
2257         BTFIXUPSET_HALF(pte_mkcleani, SRMMU_DIRTY);
2258         BTFIXUPSET_HALF(pte_mkoldi, SRMMU_REF);
2259         BTFIXUPSET_CALL(pte_mkwrite, srmmu_pte_mkwrite, BTFIXUPCALL_ORINT(SRMMU_WRITE));
2260         BTFIXUPSET_CALL(pte_mkdirty, srmmu_pte_mkdirty, BTFIXUPCALL_ORINT(SRMMU_DIRTY));
2261         BTFIXUPSET_CALL(pte_mkyoung, srmmu_pte_mkyoung, BTFIXUPCALL_ORINT(SRMMU_REF));
2262         BTFIXUPSET_CALL(update_mmu_cache, srmmu_update_mmu_cache, BTFIXUPCALL_NOP);
2263         BTFIXUPSET_CALL(destroy_context, srmmu_destroy_context, BTFIXUPCALL_NORM);
2264
2265         BTFIXUPSET_CALL(sparc_mapiorange, srmmu_mapiorange, BTFIXUPCALL_NORM);
2266         BTFIXUPSET_CALL(sparc_unmapiorange, srmmu_unmapiorange, BTFIXUPCALL_NORM);
2267
2268         BTFIXUPSET_CALL(__swp_type, srmmu_swp_type, BTFIXUPCALL_NORM);
2269         BTFIXUPSET_CALL(__swp_offset, srmmu_swp_offset, BTFIXUPCALL_NORM);
2270         BTFIXUPSET_CALL(__swp_entry, srmmu_swp_entry, BTFIXUPCALL_NORM);
2271
2272         BTFIXUPSET_CALL(mmu_info, srmmu_mmu_info, BTFIXUPCALL_NORM);
2273
2274         BTFIXUPSET_CALL(alloc_thread_info, srmmu_alloc_thread_info, BTFIXUPCALL_NORM);
2275         BTFIXUPSET_CALL(free_thread_info, srmmu_free_thread_info, BTFIXUPCALL_NORM);
2276
2277         BTFIXUPSET_CALL(pte_to_pgoff, srmmu_pte_to_pgoff, BTFIXUPCALL_NORM);
2278         BTFIXUPSET_CALL(pgoff_to_pte, srmmu_pgoff_to_pte, BTFIXUPCALL_NORM);
2279
2280         get_srmmu_type();
2281         patch_window_trap_handlers();
2282
2283 #ifdef CONFIG_SMP
2284         /* El switcheroo... */
2285
2286         BTFIXUPCOPY_CALL(local_flush_cache_all, flush_cache_all);
2287         BTFIXUPCOPY_CALL(local_flush_cache_mm, flush_cache_mm);
2288         BTFIXUPCOPY_CALL(local_flush_cache_range, flush_cache_range);
2289         BTFIXUPCOPY_CALL(local_flush_cache_page, flush_cache_page);
2290         BTFIXUPCOPY_CALL(local_flush_tlb_all, flush_tlb_all);
2291         BTFIXUPCOPY_CALL(local_flush_tlb_mm, flush_tlb_mm);
2292         BTFIXUPCOPY_CALL(local_flush_tlb_range, flush_tlb_range);
2293         BTFIXUPCOPY_CALL(local_flush_tlb_page, flush_tlb_page);
2294         BTFIXUPCOPY_CALL(local_flush_page_to_ram, __flush_page_to_ram);
2295         BTFIXUPCOPY_CALL(local_flush_sig_insns, flush_sig_insns);
2296         BTFIXUPCOPY_CALL(local_flush_page_for_dma, flush_page_for_dma);
2297
2298         BTFIXUPSET_CALL(flush_cache_all, smp_flush_cache_all, BTFIXUPCALL_NORM);
2299         BTFIXUPSET_CALL(flush_cache_mm, smp_flush_cache_mm, BTFIXUPCALL_NORM);
2300         BTFIXUPSET_CALL(flush_cache_range, smp_flush_cache_range, BTFIXUPCALL_NORM);
2301         BTFIXUPSET_CALL(flush_cache_page, smp_flush_cache_page, BTFIXUPCALL_NORM);
2302         if (sparc_cpu_model != sun4d &&
2303             sparc_cpu_model != sparc_leon) {
2304                 BTFIXUPSET_CALL(flush_tlb_all, smp_flush_tlb_all, BTFIXUPCALL_NORM);
2305                 BTFIXUPSET_CALL(flush_tlb_mm, smp_flush_tlb_mm, BTFIXUPCALL_NORM);
2306                 BTFIXUPSET_CALL(flush_tlb_range, smp_flush_tlb_range, BTFIXUPCALL_NORM);
2307                 BTFIXUPSET_CALL(flush_tlb_page, smp_flush_tlb_page, BTFIXUPCALL_NORM);
2308         }
2309         BTFIXUPSET_CALL(__flush_page_to_ram, smp_flush_page_to_ram, BTFIXUPCALL_NORM);
2310         BTFIXUPSET_CALL(flush_sig_insns, smp_flush_sig_insns, BTFIXUPCALL_NORM);
2311         BTFIXUPSET_CALL(flush_page_for_dma, smp_flush_page_for_dma, BTFIXUPCALL_NORM);
2312
2313         if (poke_srmmu == poke_viking) {
2314                 /* Avoid unnecessary cross calls. */
2315                 BTFIXUPCOPY_CALL(flush_cache_all, local_flush_cache_all);
2316                 BTFIXUPCOPY_CALL(flush_cache_mm, local_flush_cache_mm);
2317                 BTFIXUPCOPY_CALL(flush_cache_range, local_flush_cache_range);
2318                 BTFIXUPCOPY_CALL(flush_cache_page, local_flush_cache_page);
2319                 BTFIXUPCOPY_CALL(__flush_page_to_ram, local_flush_page_to_ram);
2320                 BTFIXUPCOPY_CALL(flush_sig_insns, local_flush_sig_insns);
2321                 BTFIXUPCOPY_CALL(flush_page_for_dma, local_flush_page_for_dma);
2322         }
2323 #endif
2324
2325         if (sparc_cpu_model == sun4d)
2326                 ld_mmu_iounit();
2327         else
2328                 ld_mmu_iommu();
2329 #ifdef CONFIG_SMP
2330         if (sparc_cpu_model == sun4d)
2331                 sun4d_init_smp();
2332         else if (sparc_cpu_model == sparc_leon)
2333                 leon_init_smp();
2334         else
2335                 sun4m_init_smp();
2336 #endif
2337 }