arch/sparc/mm/srmmu.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * srmmu.c:  SRMMU specific routines for memory management.
   4  *
   5  * Copyright (C) 1995 David S. Miller  (davem@caip.rutgers.edu)
   6  * Copyright (C) 1995,2002 Pete Zaitcev (zaitcev@yahoo.com)
   7  * Copyright (C) 1996 Eddie C. Dost    (ecd@skynet.be)
   8  * Copyright (C) 1997,1998 Jakub Jelinek (jj@sunsite.mff.cuni.cz)
   9  * Copyright (C) 1999,2000 Anton Blanchard (anton@samba.org)
  10  */
  11
  12 #include <linux/seq_file.h>
  13 #include <linux/spinlock.h>
  14 #include <linux/memblock.h>
  15 #include <linux/pagemap.h>
  16 #include <linux/vmalloc.h>
  17 #include <linux/kdebug.h>
  18 #include <linux/export.h>
  19 #include <linux/kernel.h>
  20 #include <linux/init.h>
  21 #include <linux/log2.h>
  22 #include <linux/gfp.h>
  23 #include <linux/fs.h>
  24 #include <linux/mm.h>
  25
  26 #include <asm/mmu_context.h>
  27 #include <asm/cacheflush.h>
  28 #include <asm/tlbflush.h>
  29 #include <asm/io-unit.h>
  30 #include <asm/pgalloc.h>
  31 #include <asm/pgtable.h>
  32 #include <asm/bitext.h>
  33 #include <asm/vaddrs.h>
  34 #include <asm/cache.h>
  35 #include <asm/traps.h>
  36 #include <asm/oplib.h>
  37 #include <asm/mbus.h>
  38 #include <asm/page.h>
  39 #include <asm/asi.h>
  40 #include <asm/smp.h>
  41 #include <asm/io.h>
  42
  43 /* Now the cpu specific definitions. */
  44 #include <asm/turbosparc.h>
  45 #include <asm/tsunami.h>
  46 #include <asm/viking.h>
  47 #include <asm/swift.h>
  48 #include <asm/leon.h>
  49 #include <asm/mxcc.h>
  50 #include <asm/ross.h>
  51
  52 #include "mm_32.h"
  53
  54 enum mbus_module srmmu_modtype;
  55 static unsigned int hwbug_bitmask;
  56 int vac_cache_size;
  57 EXPORT_SYMBOL(vac_cache_size);
  58 int vac_line_size;
  59
  60 extern struct resource sparc_iomap;
  61
  62 extern unsigned long last_valid_pfn;
  63
  64 static pgd_t *srmmu_swapper_pg_dir;
  65
  66 const struct sparc32_cachetlb_ops *sparc32_cachetlb_ops;
  67 EXPORT_SYMBOL(sparc32_cachetlb_ops);
  68
  69 #ifdef CONFIG_SMP
  70 const struct sparc32_cachetlb_ops *local_ops;
  71
  72 #define FLUSH_BEGIN(mm)
  73 #define FLUSH_END
  74 #else
  75 #define FLUSH_BEGIN(mm) if ((mm)->context != NO_CONTEXT) {
  76 #define FLUSH_END       }
  77 #endif
  78
  79 int flush_page_for_dma_global = 1;
  80
  81 char *srmmu_name;
  82
  83 ctxd_t *srmmu_ctx_table_phys;
  84 static ctxd_t *srmmu_context_table;
  85
  86 int viking_mxcc_present;
  87 static DEFINE_SPINLOCK(srmmu_context_spinlock);
  88
  89 static int is_hypersparc;
  90
  91 static int srmmu_cache_pagetables;
  92
  93 /* these will be initialized in srmmu_nocache_calcsize() */
  94 static unsigned long srmmu_nocache_size;
  95 static unsigned long srmmu_nocache_end;
  96
  97 /* 1 bit <=> 256 bytes of nocache <=> 64 PTEs */
  98 #define SRMMU_NOCACHE_BITMAP_SHIFT (PAGE_SHIFT - 4)
  99
 100 /* The context table is a nocache user with the biggest alignment needs. */
 101 #define SRMMU_NOCACHE_ALIGN_MAX (sizeof(ctxd_t)*SRMMU_MAX_CONTEXTS)
 102
 103 void *srmmu_nocache_pool;
 104 static struct bit_map srmmu_nocache_map;
 105
 106 static inline int srmmu_pmd_none(pmd_t pmd)
 107 { return !(pmd_val(pmd) & 0xFFFFFFF); }
 108
 109 /* XXX should we hyper_flush_whole_icache here - Anton */
 110 static inline void srmmu_ctxd_set(ctxd_t *ctxp, pgd_t *pgdp)
 111 {
 112         pte_t pte;
 113
 114         pte = __pte((SRMMU_ET_PTD | (__nocache_pa(pgdp) >> 4)));
 115         set_pte((pte_t *)ctxp, pte);
 116 }
 117
 118 /*
 119  * Locations of MSI Registers.
 120  */
 121 #define MSI_MBUS_ARBEN  0xe0001008      /* MBus Arbiter Enable register */
 122
 123 /*
 124  * Useful bits in the MSI Registers.
 125  */
 126 #define MSI_ASYNC_MODE  0x80000000      /* Operate the MSI asynchronously */
 127
 128 static void msi_set_sync(void)
 129 {
 130         __asm__ __volatile__ ("lda [%0] %1, %%g3\n\t"
 131                               "andn %%g3, %2, %%g3\n\t"
 132                               "sta %%g3, [%0] %1\n\t" : :
 133                               "r" (MSI_MBUS_ARBEN),
 134                               "i" (ASI_M_CTL), "r" (MSI_ASYNC_MODE) : "g3");
 135 }
 136
 137 void pmd_set(pmd_t *pmdp, pte_t *ptep)
 138 {
 139         unsigned long ptp = __nocache_pa(ptep) >> 4;
 140         set_pte((pte_t *)&pmd_val(*pmdp), __pte(SRMMU_ET_PTD | ptp));
 141 }
 142
 143 /*
 144  * size: bytes to allocate in the nocache area.
 145  * align: bytes, number to align at.
 146  * Returns the virtual address of the allocated area.
 147  */
 148 static void *__srmmu_get_nocache(int size, int align)
 149 {
 150         int offset, minsz = 1 << SRMMU_NOCACHE_BITMAP_SHIFT;
 151         unsigned long addr;
 152
 153         if (size < minsz) {
 154                 printk(KERN_ERR "Size 0x%x too small for nocache request\n",
 155                        size);
 156                 size = minsz;
 157         }
 158         if (size & (minsz - 1)) {
 159                 printk(KERN_ERR "Size 0x%x unaligned in nocache request\n",
 160                        size);
 161                 size += minsz - 1;
 162         }
 163         BUG_ON(align > SRMMU_NOCACHE_ALIGN_MAX);
 164
 165         offset = bit_map_string_get(&srmmu_nocache_map,
 166                                     size >> SRMMU_NOCACHE_BITMAP_SHIFT,
 167                                     align >> SRMMU_NOCACHE_BITMAP_SHIFT);
 168         if (offset == -1) {
 169                 printk(KERN_ERR "srmmu: out of nocache %d: %d/%d\n",
 170                        size, (int) srmmu_nocache_size,
 171                        srmmu_nocache_map.used << SRMMU_NOCACHE_BITMAP_SHIFT);
 172                 return NULL;
 173         }
 174
 175         addr = SRMMU_NOCACHE_VADDR + (offset << SRMMU_NOCACHE_BITMAP_SHIFT);
 176         return (void *)addr;
 177 }
 178
 179 void *srmmu_get_nocache(int size, int align)
 180 {
 181         void *tmp;
 182
 183         tmp = __srmmu_get_nocache(size, align);
 184
 185         if (tmp)
 186                 memset(tmp, 0, size);
 187
 188         return tmp;
 189 }
 190
 191 void srmmu_free_nocache(void *addr, int size)
 192 {
 193         unsigned long vaddr;
 194         int offset;
 195
 196         vaddr = (unsigned long)addr;
 197         if (vaddr < SRMMU_NOCACHE_VADDR) {
 198                 printk("Vaddr %lx is smaller than nocache base 0x%lx\n",
 199                     vaddr, (unsigned long)SRMMU_NOCACHE_VADDR);
 200                 BUG();
 201         }
 202         if (vaddr + size > srmmu_nocache_end) {
 203                 printk("Vaddr %lx is bigger than nocache end 0x%lx\n",
 204                     vaddr, srmmu_nocache_end);
 205                 BUG();
 206         }
 207         if (!is_power_of_2(size)) {
 208                 printk("Size 0x%x is not a power of 2\n", size);
 209                 BUG();
 210         }
 211         if (size < SRMMU_NOCACHE_BITMAP_SHIFT) {
 212                 printk("Size 0x%x is too small\n", size);
 213                 BUG();
 214         }
 215         if (vaddr & (size - 1)) {
 216                 printk("Vaddr %lx is not aligned to size 0x%x\n", vaddr, size);
 217                 BUG();
 218         }
 219
 220         offset = (vaddr - SRMMU_NOCACHE_VADDR) >> SRMMU_NOCACHE_BITMAP_SHIFT;
 221         size = size >> SRMMU_NOCACHE_BITMAP_SHIFT;
 222
 223         bit_map_clear(&srmmu_nocache_map, offset, size);
 224 }
 225
 226 static void srmmu_early_allocate_ptable_skeleton(unsigned long start,
 227                                                  unsigned long end);
 228
 229 /* Return how much physical memory we have.  */
 230 static unsigned long __init probe_memory(void)
 231 {
 232         unsigned long total = 0;
 233         int i;
 234
 235         for (i = 0; sp_banks[i].num_bytes; i++)
 236                 total += sp_banks[i].num_bytes;
 237
 238         return total;
 239 }
 240
 241 /*
 242  * Reserve nocache dynamically proportionally to the amount of
 243  * system RAM. -- Tomas Szepe <szepe@pinerecords.com>, June 2002
 244  */
 245 static void __init srmmu_nocache_calcsize(void)
 246 {
 247         unsigned long sysmemavail = probe_memory() / 1024;
 248         int srmmu_nocache_npages;
 249
 250         srmmu_nocache_npages =
 251                 sysmemavail / SRMMU_NOCACHE_ALCRATIO / 1024 * 256;
 252
 253  /* P3 XXX The 4x overuse: corroborated by /proc/meminfo. */
 254         // if (srmmu_nocache_npages < 256) srmmu_nocache_npages = 256;
 255         if (srmmu_nocache_npages < SRMMU_MIN_NOCACHE_PAGES)
 256                 srmmu_nocache_npages = SRMMU_MIN_NOCACHE_PAGES;
 257
 258         /* anything above 1280 blows up */
 259         if (srmmu_nocache_npages > SRMMU_MAX_NOCACHE_PAGES)
 260                 srmmu_nocache_npages = SRMMU_MAX_NOCACHE_PAGES;
 261
 262         srmmu_nocache_size = srmmu_nocache_npages * PAGE_SIZE;
 263         srmmu_nocache_end = SRMMU_NOCACHE_VADDR + srmmu_nocache_size;
 264 }
 265
 266 static void __init srmmu_nocache_init(void)
 267 {
 268         void *srmmu_nocache_bitmap;
 269         unsigned int bitmap_bits;
 270         pgd_t *pgd;
 271         p4d_t *p4d;
 272         pud_t *pud;
 273         pmd_t *pmd;
 274         pte_t *pte;
 275         unsigned long paddr, vaddr;
 276         unsigned long pteval;
 277
 278         bitmap_bits = srmmu_nocache_size >> SRMMU_NOCACHE_BITMAP_SHIFT;
 279
 280         srmmu_nocache_pool = memblock_alloc(srmmu_nocache_size,
 281                                             SRMMU_NOCACHE_ALIGN_MAX);
 282         if (!srmmu_nocache_pool)
 283                 panic("%s: Failed to allocate %lu bytes align=0x%x\n",
 284                       __func__, srmmu_nocache_size, SRMMU_NOCACHE_ALIGN_MAX);
 285         memset(srmmu_nocache_pool, 0, srmmu_nocache_size);
 286
 287         srmmu_nocache_bitmap =
 288                 memblock_alloc(BITS_TO_LONGS(bitmap_bits) * sizeof(long),
 289                                SMP_CACHE_BYTES);
 290         if (!srmmu_nocache_bitmap)
 291                 panic("%s: Failed to allocate %zu bytes\n", __func__,
 292                       BITS_TO_LONGS(bitmap_bits) * sizeof(long));
 293         bit_map_init(&srmmu_nocache_map, srmmu_nocache_bitmap, bitmap_bits);
 294
 295         srmmu_swapper_pg_dir = __srmmu_get_nocache(SRMMU_PGD_TABLE_SIZE, SRMMU_PGD_TABLE_SIZE);
 296         memset(__nocache_fix(srmmu_swapper_pg_dir), 0, SRMMU_PGD_TABLE_SIZE);
 297         init_mm.pgd = srmmu_swapper_pg_dir;
 298
 299         srmmu_early_allocate_ptable_skeleton(SRMMU_NOCACHE_VADDR, srmmu_nocache_end);
 300
 301         paddr = __pa((unsigned long)srmmu_nocache_pool);
 302         vaddr = SRMMU_NOCACHE_VADDR;
 303
 304         while (vaddr < srmmu_nocache_end) {
 305                 pgd = pgd_offset_k(vaddr);
 306                 p4d = p4d_offset(pgd, vaddr);
 307                 pud = pud_offset(p4d, vaddr);
 308                 pmd = pmd_offset(__nocache_fix(pud), vaddr);
 309                 pte = pte_offset_kernel(__nocache_fix(pmd), vaddr);
 310
 311                 pteval = ((paddr >> 4) | SRMMU_ET_PTE | SRMMU_PRIV);
 312
 313                 if (srmmu_cache_pagetables)
 314                         pteval |= SRMMU_CACHE;
 315
 316                 set_pte(__nocache_fix(pte), __pte(pteval));
 317
 318                 vaddr += PAGE_SIZE;
 319                 paddr += PAGE_SIZE;
 320         }
 321
 322         flush_cache_all();
 323         flush_tlb_all();
 324 }
 325
 326 pgd_t *get_pgd_fast(void)
 327 {
 328         pgd_t *pgd = NULL;
 329
 330         pgd = __srmmu_get_nocache(SRMMU_PGD_TABLE_SIZE, SRMMU_PGD_TABLE_SIZE);
 331         if (pgd) {
 332                 pgd_t *init = pgd_offset_k(0);
 333                 memset(pgd, 0, USER_PTRS_PER_PGD * sizeof(pgd_t));
 334                 memcpy(pgd + USER_PTRS_PER_PGD, init + USER_PTRS_PER_PGD,
 335                                                 (PTRS_PER_PGD - USER_PTRS_PER_PGD) * sizeof(pgd_t));
 336         }
 337
 338         return pgd;
 339 }
 340
 341 /*
 342  * Hardware needs alignment to 256 only, but we align to whole page size
 343  * to reduce fragmentation problems due to the buddy principle.
 344  * XXX Provide actual fragmentation statistics in /proc.
 345  *
 346  * Alignments up to the page size are the same for physical and virtual
 347  * addresses of the nocache area.
 348  */
 349 pgtable_t pte_alloc_one(struct mm_struct *mm)
 350 {
 351         pte_t *ptep;
 352         struct page *page;
 353
 354         if ((ptep = pte_alloc_one_kernel(mm)) == 0)
 355                 return NULL;
 356         page = pfn_to_page(__nocache_pa((unsigned long)ptep) >> PAGE_SHIFT);
 357         spin_lock(&mm->page_table_lock);
 358         if (page_ref_inc_return(page) == 2 && !pgtable_pte_page_ctor(page)) {
 359                 page_ref_dec(page);
 360                 ptep = NULL;
 361         }
 362         spin_unlock(&mm->page_table_lock);
 363
 364         return ptep;
 365 }
 366
 367 void pte_free(struct mm_struct *mm, pgtable_t ptep)
 368 {
 369         struct page *page;
 370
 371         page = pfn_to_page(__nocache_pa((unsigned long)ptep) >> PAGE_SHIFT);
 372         spin_lock(&mm->page_table_lock);
 373         if (page_ref_dec_return(page) == 1)
 374                 pgtable_pte_page_dtor(page);
 375         spin_unlock(&mm->page_table_lock);
 376
 377         srmmu_free_nocache(ptep, SRMMU_PTE_TABLE_SIZE);
 378 }
 379
 380 /* context handling - a dynamically sized pool is used */
 381 #define NO_CONTEXT      -1
 382
 383 struct ctx_list {
 384         struct ctx_list *next;
 385         struct ctx_list *prev;
 386         unsigned int ctx_number;
 387         struct mm_struct *ctx_mm;
 388 };
 389
 390 static struct ctx_list *ctx_list_pool;
 391 static struct ctx_list ctx_free;
 392 static struct ctx_list ctx_used;
 393
 394 /* At boot time we determine the number of contexts */
 395 static int num_contexts;
 396
 397 static inline void remove_from_ctx_list(struct ctx_list *entry)
 398 {
 399         entry->next->prev = entry->prev;
 400         entry->prev->next = entry->next;
 401 }
 402
 403 static inline void add_to_ctx_list(struct ctx_list *head, struct ctx_list *entry)
 404 {
 405         entry->next = head;
 406         (entry->prev = head->prev)->next = entry;
 407         head->prev = entry;
 408 }
 409 #define add_to_free_ctxlist(entry) add_to_ctx_list(&ctx_free, entry)
 410 #define add_to_used_ctxlist(entry) add_to_ctx_list(&ctx_used, entry)
 411
 412
 413 static inline void alloc_context(struct mm_struct *old_mm, struct mm_struct *mm)
 414 {
 415         struct ctx_list *ctxp;
 416
 417         ctxp = ctx_free.next;
 418         if (ctxp != &ctx_free) {
 419                 remove_from_ctx_list(ctxp);
 420                 add_to_used_ctxlist(ctxp);
 421                 mm->context = ctxp->ctx_number;
 422                 ctxp->ctx_mm = mm;
 423                 return;
 424         }
 425         ctxp = ctx_used.next;
 426         if (ctxp->ctx_mm == old_mm)
 427                 ctxp = ctxp->next;
 428         if (ctxp == &ctx_used)
 429                 panic("out of mmu contexts");
 430         flush_cache_mm(ctxp->ctx_mm);
 431         flush_tlb_mm(ctxp->ctx_mm);
 432         remove_from_ctx_list(ctxp);
 433         add_to_used_ctxlist(ctxp);
 434         ctxp->ctx_mm->context = NO_CONTEXT;
 435         ctxp->ctx_mm = mm;
 436         mm->context = ctxp->ctx_number;
 437 }
 438
 439 static inline void free_context(int context)
 440 {
 441         struct ctx_list *ctx_old;
 442
 443         ctx_old = ctx_list_pool + context;
 444         remove_from_ctx_list(ctx_old);
 445         add_to_free_ctxlist(ctx_old);
 446 }
 447
 448 static void __init sparc_context_init(int numctx)
 449 {
 450         int ctx;
 451         unsigned long size;
 452
 453         size = numctx * sizeof(struct ctx_list);
 454         ctx_list_pool = memblock_alloc(size, SMP_CACHE_BYTES);
 455         if (!ctx_list_pool)
 456                 panic("%s: Failed to allocate %lu bytes\n", __func__, size);
 457
 458         for (ctx = 0; ctx < numctx; ctx++) {
 459                 struct ctx_list *clist;
 460
 461                 clist = (ctx_list_pool + ctx);
 462                 clist->ctx_number = ctx;
 463                 clist->ctx_mm = NULL;
 464         }
 465         ctx_free.next = ctx_free.prev = &ctx_free;
 466         ctx_used.next = ctx_used.prev = &ctx_used;
 467         for (ctx = 0; ctx < numctx; ctx++)
 468                 add_to_free_ctxlist(ctx_list_pool + ctx);
 469 }
 470
 471 void switch_mm(struct mm_struct *old_mm, struct mm_struct *mm,
 472                struct task_struct *tsk)
 473 {
 474         unsigned long flags;
 475
 476         if (mm->context == NO_CONTEXT) {
 477                 spin_lock_irqsave(&srmmu_context_spinlock, flags);
 478                 alloc_context(old_mm, mm);
 479                 spin_unlock_irqrestore(&srmmu_context_spinlock, flags);
 480                 srmmu_ctxd_set(&srmmu_context_table[mm->context], mm->pgd);
 481         }
 482
 483         if (sparc_cpu_model == sparc_leon)
 484                 leon_switch_mm();
 485
 486         if (is_hypersparc)
 487                 hyper_flush_whole_icache();
 488
 489         srmmu_set_context(mm->context);
 490 }
 491
 492 /* Low level IO area allocation on the SRMMU. */
 493 static inline void srmmu_mapioaddr(unsigned long physaddr,
 494                                    unsigned long virt_addr, int bus_type)
 495 {
 496         pgd_t *pgdp;
 497         p4d_t *p4dp;
 498         pud_t *pudp;
 499         pmd_t *pmdp;
 500         pte_t *ptep;
 501         unsigned long tmp;
 502
 503         physaddr &= PAGE_MASK;
 504         pgdp = pgd_offset_k(virt_addr);
 505         p4dp = p4d_offset(pgdp, virt_addr);
 506         pudp = pud_offset(p4dp, virt_addr);
 507         pmdp = pmd_offset(pudp, virt_addr);
 508         ptep = pte_offset_kernel(pmdp, virt_addr);
 509         tmp = (physaddr >> 4) | SRMMU_ET_PTE;
 510
 511         /* I need to test whether this is consistent over all
 512          * sun4m's.  The bus_type represents the upper 4 bits of
 513          * 36-bit physical address on the I/O space lines...
 514          */
 515         tmp |= (bus_type << 28);
 516         tmp |= SRMMU_PRIV;
 517         __flush_page_to_ram(virt_addr);
 518         set_pte(ptep, __pte(tmp));
 519 }
 520
 521 void srmmu_mapiorange(unsigned int bus, unsigned long xpa,
 522                       unsigned long xva, unsigned int len)
 523 {
 524         while (len != 0) {
 525                 len -= PAGE_SIZE;
 526                 srmmu_mapioaddr(xpa, xva, bus);
 527                 xva += PAGE_SIZE;
 528                 xpa += PAGE_SIZE;
 529         }
 530         flush_tlb_all();
 531 }
 532
 533 static inline void srmmu_unmapioaddr(unsigned long virt_addr)
 534 {
 535         pgd_t *pgdp;
 536         p4d_t *p4dp;
 537         pud_t *pudp;
 538         pmd_t *pmdp;
 539         pte_t *ptep;
 540
 541
 542         pgdp = pgd_offset_k(virt_addr);
 543         p4dp = p4d_offset(pgdp, virt_addr);
 544         pudp = pud_offset(p4dp, virt_addr);
 545         pmdp = pmd_offset(pudp, virt_addr);
 546         ptep = pte_offset_kernel(pmdp, virt_addr);
 547
 548         /* No need to flush uncacheable page. */
 549         __pte_clear(ptep);
 550 }
 551
 552 void srmmu_unmapiorange(unsigned long virt_addr, unsigned int len)
 553 {
 554         while (len != 0) {
 555                 len -= PAGE_SIZE;
 556                 srmmu_unmapioaddr(virt_addr);
 557                 virt_addr += PAGE_SIZE;
 558         }
 559         flush_tlb_all();
 560 }
 561
 562 /* tsunami.S */
 563 extern void tsunami_flush_cache_all(void);
 564 extern void tsunami_flush_cache_mm(struct mm_struct *mm);
 565 extern void tsunami_flush_cache_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
 566 extern void tsunami_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
 567 extern void tsunami_flush_page_to_ram(unsigned long page);
 568 extern void tsunami_flush_page_for_dma(unsigned long page);
 569 extern void tsunami_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr);
 570 extern void tsunami_flush_tlb_all(void);
 571 extern void tsunami_flush_tlb_mm(struct mm_struct *mm);
 572 extern void tsunami_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
 573 extern void tsunami_flush_tlb_page(struct vm_area_struct *vma, unsigned long page);
 574 extern void tsunami_setup_blockops(void);
 575
 576 /* swift.S */
 577 extern void swift_flush_cache_all(void);
 578 extern void swift_flush_cache_mm(struct mm_struct *mm);
 579 extern void swift_flush_cache_range(struct vm_area_struct *vma,
 580                                     unsigned long start, unsigned long end);
 581 extern void swift_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
 582 extern void swift_flush_page_to_ram(unsigned long page);
 583 extern void swift_flush_page_for_dma(unsigned long page);
 584 extern void swift_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr);
 585 extern void swift_flush_tlb_all(void);
 586 extern void swift_flush_tlb_mm(struct mm_struct *mm);
 587 extern void swift_flush_tlb_range(struct vm_area_struct *vma,
 588                                   unsigned long start, unsigned long end);
 589 extern void swift_flush_tlb_page(struct vm_area_struct *vma, unsigned long page);
 590
 591 #if 0  /* P3: deadwood to debug precise flushes on Swift. */
 592 void swift_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
 593 {
 594         int cctx, ctx1;
 595
 596         page &= PAGE_MASK;
 597         if ((ctx1 = vma->vm_mm->context) != -1) {
 598                 cctx = srmmu_get_context();
 599 /* Is context # ever different from current context? P3 */
 600                 if (cctx != ctx1) {
 601                         printk("flush ctx %02x curr %02x\n", ctx1, cctx);
 602                         srmmu_set_context(ctx1);
 603                         swift_flush_page(page);
 604                         __asm__ __volatile__("sta %%g0, [%0] %1\n\t" : :
 605                                         "r" (page), "i" (ASI_M_FLUSH_PROBE));
 606                         srmmu_set_context(cctx);
 607                 } else {
 608                          /* Rm. prot. bits from virt. c. */
 609                         /* swift_flush_cache_all(); */
 610                         /* swift_flush_cache_page(vma, page); */
 611                         swift_flush_page(page);
 612
 613                         __asm__ __volatile__("sta %%g0, [%0] %1\n\t" : :
 614                                 "r" (page), "i" (ASI_M_FLUSH_PROBE));
 615                         /* same as above: srmmu_flush_tlb_page() */
 616                 }
 617         }
 618 }
 619 #endif
 620
 621 /*
 622  * The following are all MBUS based SRMMU modules, and therefore could
 623  * be found in a multiprocessor configuration.  On the whole, these
 624  * chips seems to be much more touchy about DVMA and page tables
 625  * with respect to cache coherency.
 626  */
 627
 628 /* viking.S */
 629 extern void viking_flush_cache_all(void);
 630 extern void viking_flush_cache_mm(struct mm_struct *mm);
 631 extern void viking_flush_cache_range(struct vm_area_struct *vma, unsigned long start,
 632                                      unsigned long end);
 633 extern void viking_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
 634 extern void viking_flush_page_to_ram(unsigned long page);
 635 extern void viking_flush_page_for_dma(unsigned long page);
 636 extern void viking_flush_sig_insns(struct mm_struct *mm, unsigned long addr);
 637 extern void viking_flush_page(unsigned long page);
 638 extern void viking_mxcc_flush_page(unsigned long page);
 639 extern void viking_flush_tlb_all(void);
 640 extern void viking_flush_tlb_mm(struct mm_struct *mm);
 641 extern void viking_flush_tlb_range(struct vm_area_struct *vma, unsigned long start,
 642                                    unsigned long end);
 643 extern void viking_flush_tlb_page(struct vm_area_struct *vma,
 644                                   unsigned long page);
 645 extern void sun4dsmp_flush_tlb_all(void);
 646 extern void sun4dsmp_flush_tlb_mm(struct mm_struct *mm);
 647 extern void sun4dsmp_flush_tlb_range(struct vm_area_struct *vma, unsigned long start,
 648                                    unsigned long end);
 649 extern void sun4dsmp_flush_tlb_page(struct vm_area_struct *vma,
 650                                   unsigned long page);
 651
 652 /* hypersparc.S */
 653 extern void hypersparc_flush_cache_all(void);
 654 extern void hypersparc_flush_cache_mm(struct mm_struct *mm);
 655 extern void hypersparc_flush_cache_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
 656 extern void hypersparc_flush_cache_page(struct vm_area_struct *vma, unsigned long page);
 657 extern void hypersparc_flush_page_to_ram(unsigned long page);
 658 extern void hypersparc_flush_page_for_dma(unsigned long page);
 659 extern void hypersparc_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr);
 660 extern void hypersparc_flush_tlb_all(void);
 661 extern void hypersparc_flush_tlb_mm(struct mm_struct *mm);
 662 extern void hypersparc_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end);
 663 extern void hypersparc_flush_tlb_page(struct vm_area_struct *vma, unsigned long page);
 664 extern void hypersparc_setup_blockops(void);
 665
 666 /*
 667  * NOTE: All of this startup code assumes the low 16mb (approx.) of
 668  *       kernel mappings are done with one single contiguous chunk of
 669  *       ram.  On small ram machines (classics mainly) we only get
 670  *       around 8mb mapped for us.
 671  */
 672
 673 static void __init early_pgtable_allocfail(char *type)
 674 {
 675         prom_printf("inherit_prom_mappings: Cannot alloc kernel %s.\n", type);
 676         prom_halt();
 677 }
 678
 679 static void __init srmmu_early_allocate_ptable_skeleton(unsigned long start,
 680                                                         unsigned long end)
 681 {
 682         pgd_t *pgdp;
 683         p4d_t *p4dp;
 684         pud_t *pudp;
 685         pmd_t *pmdp;
 686         pte_t *ptep;
 687
 688         while (start < end) {
 689                 pgdp = pgd_offset_k(start);
 690                 p4dp = p4d_offset(pgdp, start);
 691                 pudp = pud_offset(p4dp, start);
 692                 if (pud_none(*(pud_t *)__nocache_fix(pudp))) {
 693                         pmdp = __srmmu_get_nocache(
 694                             SRMMU_PMD_TABLE_SIZE, SRMMU_PMD_TABLE_SIZE);
 695                         if (pmdp == NULL)
 696                                 early_pgtable_allocfail("pmd");
 697                         memset(__nocache_fix(pmdp), 0, SRMMU_PMD_TABLE_SIZE);
 698                         pud_set(__nocache_fix(pudp), pmdp);
 699                 }
 700                 pmdp = pmd_offset(__nocache_fix(pudp), start);
 701                 if (srmmu_pmd_none(*(pmd_t *)__nocache_fix(pmdp))) {
 702                         ptep = __srmmu_get_nocache(PTE_SIZE, PTE_SIZE);
 703                         if (ptep == NULL)
 704                                 early_pgtable_allocfail("pte");
 705                         memset(__nocache_fix(ptep), 0, PTE_SIZE);
 706                         pmd_set(__nocache_fix(pmdp), ptep);
 707                 }
 708                 if (start > (0xffffffffUL - PMD_SIZE))
 709                         break;
 710                 start = (start + PMD_SIZE) & PMD_MASK;
 711         }
 712 }
 713
 714 static void __init srmmu_allocate_ptable_skeleton(unsigned long start,
 715                                                   unsigned long end)
 716 {
 717         pgd_t *pgdp;
 718         p4d_t *p4dp;
 719         pud_t *pudp;
 720         pmd_t *pmdp;
 721         pte_t *ptep;
 722
 723         while (start < end) {
 724                 pgdp = pgd_offset_k(start);
 725                 p4dp = p4d_offset(pgdp, start);
 726                 pudp = pud_offset(p4dp, start);
 727                 if (pud_none(*pudp)) {
 728                         pmdp = __srmmu_get_nocache(SRMMU_PMD_TABLE_SIZE, SRMMU_PMD_TABLE_SIZE);
 729                         if (pmdp == NULL)
 730                                 early_pgtable_allocfail("pmd");
 731                         memset(pmdp, 0, SRMMU_PMD_TABLE_SIZE);
 732                         pud_set((pud_t *)pgdp, pmdp);
 733                 }
 734                 pmdp = pmd_offset(pudp, start);
 735                 if (srmmu_pmd_none(*pmdp)) {
 736                         ptep = __srmmu_get_nocache(PTE_SIZE,
 737                                                              PTE_SIZE);
 738                         if (ptep == NULL)
 739                                 early_pgtable_allocfail("pte");
 740                         memset(ptep, 0, PTE_SIZE);
 741                         pmd_set(pmdp, ptep);
 742                 }
 743                 if (start > (0xffffffffUL - PMD_SIZE))
 744                         break;
 745                 start = (start + PMD_SIZE) & PMD_MASK;
 746         }
 747 }
 748
 749 /* These flush types are not available on all chips... */
 750 static inline unsigned long srmmu_probe(unsigned long vaddr)
 751 {
 752         unsigned long retval;
 753
 754         if (sparc_cpu_model != sparc_leon) {
 755
 756                 vaddr &= PAGE_MASK;
 757                 __asm__ __volatile__("lda [%1] %2, %0\n\t" :
 758                                      "=r" (retval) :
 759                                      "r" (vaddr | 0x400), "i" (ASI_M_FLUSH_PROBE));
 760         } else {
 761                 retval = leon_swprobe(vaddr, NULL);
 762         }
 763         return retval;
 764 }
 765
 766 /*
 767  * This is much cleaner than poking around physical address space
 768  * looking at the prom's page table directly which is what most
 769  * other OS's do.  Yuck... this is much better.
 770  */
 771 static void __init srmmu_inherit_prom_mappings(unsigned long start,
 772                                                unsigned long end)
 773 {
 774         unsigned long probed;
 775         unsigned long addr;
 776         pgd_t *pgdp;
 777         p4d_t *p4dp;
 778         pud_t *pudp;
 779         pmd_t *pmdp;
 780         pte_t *ptep;
 781         int what; /* 0 = normal-pte, 1 = pmd-level pte, 2 = pgd-level pte */
 782
 783         while (start <= end) {
 784                 if (start == 0)
 785                         break; /* probably wrap around */
 786                 if (start == 0xfef00000)
 787                         start = KADB_DEBUGGER_BEGVM;
 788                 probed = srmmu_probe(start);
 789                 if (!probed) {
 790                         /* continue probing until we find an entry */
 791                         start += PAGE_SIZE;
 792                         continue;
 793                 }
 794
 795                 /* A red snapper, see what it really is. */
 796                 what = 0;
 797                 addr = start - PAGE_SIZE;
 798
 799                 if (!(start & ~(PMD_MASK))) {
 800                         if (srmmu_probe(addr + PMD_SIZE) == probed)
 801                                 what = 1;
 802                 }
 803
 804                 if (!(start & ~(PGDIR_MASK))) {
 805                         if (srmmu_probe(addr + PGDIR_SIZE) == probed)
 806                                 what = 2;
 807                 }
 808
 809                 pgdp = pgd_offset_k(start);
 810                 p4dp = p4d_offset(pgdp, start);
 811                 pudp = pud_offset(p4dp, start);
 812                 if (what == 2) {
 813                         *(pgd_t *)__nocache_fix(pgdp) = __pgd(probed);
 814                         start += PGDIR_SIZE;
 815                         continue;
 816                 }
 817                 if (pud_none(*(pud_t *)__nocache_fix(pudp))) {
 818                         pmdp = __srmmu_get_nocache(SRMMU_PMD_TABLE_SIZE,
 819                                                    SRMMU_PMD_TABLE_SIZE);
 820                         if (pmdp == NULL)
 821                                 early_pgtable_allocfail("pmd");
 822                         memset(__nocache_fix(pmdp), 0, SRMMU_PMD_TABLE_SIZE);
 823                         pud_set(__nocache_fix(pudp), pmdp);
 824                 }
 825                 pmdp = pmd_offset(__nocache_fix(pudp), start);
 826                 if (what == 1) {
 827                         *(pmd_t *)__nocache_fix(pmdp) = __pmd(probed);
 828                         start += PMD_SIZE;
 829                         continue;
 830                 }
 831                 if (srmmu_pmd_none(*(pmd_t *)__nocache_fix(pmdp))) {
 832                         ptep = __srmmu_get_nocache(PTE_SIZE, PTE_SIZE);
 833                         if (ptep == NULL)
 834                                 early_pgtable_allocfail("pte");
 835                         memset(__nocache_fix(ptep), 0, PTE_SIZE);
 836                         pmd_set(__nocache_fix(pmdp), ptep);
 837                 }
 838                 ptep = pte_offset_kernel(__nocache_fix(pmdp), start);
 839                 *(pte_t *)__nocache_fix(ptep) = __pte(probed);
 840                 start += PAGE_SIZE;
 841         }
 842 }
 843
 844 #define KERNEL_PTE(page_shifted) ((page_shifted)|SRMMU_CACHE|SRMMU_PRIV|SRMMU_VALID)
 845
 846 /* Create a third-level SRMMU 16MB page mapping. */
 847 static void __init do_large_mapping(unsigned long vaddr, unsigned long phys_base)
 848 {
 849         pgd_t *pgdp = pgd_offset_k(vaddr);
 850         unsigned long big_pte;
 851
 852         big_pte = KERNEL_PTE(phys_base >> 4);
 853         *(pgd_t *)__nocache_fix(pgdp) = __pgd(big_pte);
 854 }
 855
 856 /* Map sp_bank entry SP_ENTRY, starting at virtual address VBASE. */
 857 static unsigned long __init map_spbank(unsigned long vbase, int sp_entry)
 858 {
 859         unsigned long pstart = (sp_banks[sp_entry].base_addr & PGDIR_MASK);
 860         unsigned long vstart = (vbase & PGDIR_MASK);
 861         unsigned long vend = PGDIR_ALIGN(vbase + sp_banks[sp_entry].num_bytes);
 862         /* Map "low" memory only */
 863         const unsigned long min_vaddr = PAGE_OFFSET;
 864         const unsigned long max_vaddr = PAGE_OFFSET + SRMMU_MAXMEM;
 865
 866         if (vstart < min_vaddr || vstart >= max_vaddr)
 867                 return vstart;
 868
 869         if (vend > max_vaddr || vend < min_vaddr)
 870                 vend = max_vaddr;
 871
 872         while (vstart < vend) {
 873                 do_large_mapping(vstart, pstart);
 874                 vstart += PGDIR_SIZE; pstart += PGDIR_SIZE;
 875         }
 876         return vstart;
 877 }
 878
 879 static void __init map_kernel(void)
 880 {
 881         int i;
 882
 883         if (phys_base > 0) {
 884                 do_large_mapping(PAGE_OFFSET, phys_base);
 885         }
 886
 887         for (i = 0; sp_banks[i].num_bytes != 0; i++) {
 888                 map_spbank((unsigned long)__va(sp_banks[i].base_addr), i);
 889         }
 890 }
 891
 892 void (*poke_srmmu)(void) = NULL;
 893
 894 void __init srmmu_paging_init(void)
 895 {
 896         int i;
 897         phandle cpunode;
 898         char node_str[128];
 899         pgd_t *pgd;
 900         p4d_t *p4d;
 901         pud_t *pud;
 902         pmd_t *pmd;
 903         pte_t *pte;
 904         unsigned long pages_avail;
 905
 906         init_mm.context = (unsigned long) NO_CONTEXT;
 907         sparc_iomap.start = SUN4M_IOBASE_VADDR; /* 16MB of IOSPACE on all sun4m's. */
 908
 909         if (sparc_cpu_model == sun4d)
 910                 num_contexts = 65536; /* We know it is Viking */
 911         else {
 912                 /* Find the number of contexts on the srmmu. */
 913                 cpunode = prom_getchild(prom_root_node);
 914                 num_contexts = 0;
 915                 while (cpunode != 0) {
 916                         prom_getstring(cpunode, "device_type", node_str, sizeof(node_str));
 917                         if (!strcmp(node_str, "cpu")) {
 918                                 num_contexts = prom_getintdefault(cpunode, "mmu-nctx", 0x8);
 919                                 break;
 920                         }
 921                         cpunode = prom_getsibling(cpunode);
 922                 }
 923         }
 924
 925         if (!num_contexts) {
 926                 prom_printf("Something wrong, can't find cpu node in paging_init.\n");
 927                 prom_halt();
 928         }
 929
 930         pages_avail = 0;
 931         last_valid_pfn = bootmem_init(&pages_avail);
 932
 933         srmmu_nocache_calcsize();
 934         srmmu_nocache_init();
 935         srmmu_inherit_prom_mappings(0xfe400000, (LINUX_OPPROM_ENDVM - PAGE_SIZE));
 936         map_kernel();
 937
 938         /* ctx table has to be physically aligned to its size */
 939         srmmu_context_table = __srmmu_get_nocache(num_contexts * sizeof(ctxd_t), num_contexts * sizeof(ctxd_t));
 940         srmmu_ctx_table_phys = (ctxd_t *)__nocache_pa(srmmu_context_table);
 941
 942         for (i = 0; i < num_contexts; i++)
 943                 srmmu_ctxd_set((ctxd_t *)__nocache_fix(&srmmu_context_table[i]), srmmu_swapper_pg_dir);
 944
 945         flush_cache_all();
 946         srmmu_set_ctable_ptr((unsigned long)srmmu_ctx_table_phys);
 947 #ifdef CONFIG_SMP
 948         /* Stop from hanging here... */
 949         local_ops->tlb_all();
 950 #else
 951         flush_tlb_all();
 952 #endif
 953         poke_srmmu();
 954
 955         srmmu_allocate_ptable_skeleton(sparc_iomap.start, IOBASE_END);
 956         srmmu_allocate_ptable_skeleton(DVMA_VADDR, DVMA_END);
 957
 958         srmmu_allocate_ptable_skeleton(
 959                 __fix_to_virt(__end_of_fixed_addresses - 1), FIXADDR_TOP);
 960         srmmu_allocate_ptable_skeleton(PKMAP_BASE, PKMAP_END);
 961
 962         pgd = pgd_offset_k(PKMAP_BASE);
 963         p4d = p4d_offset(pgd, PKMAP_BASE);
 964         pud = pud_offset(p4d, PKMAP_BASE);
 965         pmd = pmd_offset(pud, PKMAP_BASE);
 966         pte = pte_offset_kernel(pmd, PKMAP_BASE);
 967         pkmap_page_table = pte;
 968
 969         flush_cache_all();
 970         flush_tlb_all();
 971
 972         sparc_context_init(num_contexts);
 973
 974         kmap_init();
 975
 976         {
 977                 unsigned long max_zone_pfn[MAX_NR_ZONES] = { 0 };
 978
 979                 max_zone_pfn[ZONE_DMA] = max_low_pfn;
 980                 max_zone_pfn[ZONE_NORMAL] = max_low_pfn;
 981                 max_zone_pfn[ZONE_HIGHMEM] = highend_pfn;
 982
 983                 free_area_init(max_zone_pfn);
 984         }
 985 }
 986
 987 void mmu_info(struct seq_file *m)
 988 {
 989         seq_printf(m,
 990                    "MMU type\t: %s\n"
 991                    "contexts\t: %d\n"
 992                    "nocache total\t: %ld\n"
 993                    "nocache used\t: %d\n",
 994                    srmmu_name,
 995                    num_contexts,
 996                    srmmu_nocache_size,
 997                    srmmu_nocache_map.used << SRMMU_NOCACHE_BITMAP_SHIFT);
 998 }
 999
1000 int init_new_context(struct task_struct *tsk, struct mm_struct *mm)
1001 {
1002         mm->context = NO_CONTEXT;
1003         return 0;
1004 }
1005
1006 void destroy_context(struct mm_struct *mm)
1007 {
1008         unsigned long flags;
1009
1010         if (mm->context != NO_CONTEXT) {
1011                 flush_cache_mm(mm);
1012                 srmmu_ctxd_set(&srmmu_context_table[mm->context], srmmu_swapper_pg_dir);
1013                 flush_tlb_mm(mm);
1014                 spin_lock_irqsave(&srmmu_context_spinlock, flags);
1015                 free_context(mm->context);
1016                 spin_unlock_irqrestore(&srmmu_context_spinlock, flags);
1017                 mm->context = NO_CONTEXT;
1018         }
1019 }
1020
1021 /* Init various srmmu chip types. */
1022 static void __init srmmu_is_bad(void)
1023 {
1024         prom_printf("Could not determine SRMMU chip type.\n");
1025         prom_halt();
1026 }
1027
1028 static void __init init_vac_layout(void)
1029 {
1030         phandle nd;
1031         int cache_lines;
1032         char node_str[128];
1033 #ifdef CONFIG_SMP
1034         int cpu = 0;
1035         unsigned long max_size = 0;
1036         unsigned long min_line_size = 0x10000000;
1037 #endif
1038
1039         nd = prom_getchild(prom_root_node);
1040         while ((nd = prom_getsibling(nd)) != 0) {
1041                 prom_getstring(nd, "device_type", node_str, sizeof(node_str));
1042                 if (!strcmp(node_str, "cpu")) {
1043                         vac_line_size = prom_getint(nd, "cache-line-size");
1044                         if (vac_line_size == -1) {
1045                                 prom_printf("can't determine cache-line-size, halting.\n");
1046                                 prom_halt();
1047                         }
1048                         cache_lines = prom_getint(nd, "cache-nlines");
1049                         if (cache_lines == -1) {
1050                                 prom_printf("can't determine cache-nlines, halting.\n");
1051                                 prom_halt();
1052                         }
1053
1054                         vac_cache_size = cache_lines * vac_line_size;
1055 #ifdef CONFIG_SMP
1056                         if (vac_cache_size > max_size)
1057                                 max_size = vac_cache_size;
1058                         if (vac_line_size < min_line_size)
1059                                 min_line_size = vac_line_size;
1060                         //FIXME: cpus not contiguous!!
1061                         cpu++;
1062                         if (cpu >= nr_cpu_ids || !cpu_online(cpu))
1063                                 break;
1064 #else
1065                         break;
1066 #endif
1067                 }
1068         }
1069         if (nd == 0) {
1070                 prom_printf("No CPU nodes found, halting.\n");
1071                 prom_halt();
1072         }
1073 #ifdef CONFIG_SMP
1074         vac_cache_size = max_size;
1075         vac_line_size = min_line_size;
1076 #endif
1077         printk("SRMMU: Using VAC size of %d bytes, line size %d bytes.\n",
1078                (int)vac_cache_size, (int)vac_line_size);
1079 }
1080
1081 static void poke_hypersparc(void)
1082 {
1083         volatile unsigned long clear;
1084         unsigned long mreg = srmmu_get_mmureg();
1085
1086         hyper_flush_unconditional_combined();
1087
1088         mreg &= ~(HYPERSPARC_CWENABLE);
1089         mreg |= (HYPERSPARC_CENABLE | HYPERSPARC_WBENABLE);
1090         mreg |= (HYPERSPARC_CMODE);
1091
1092         srmmu_set_mmureg(mreg);
1093
1094 #if 0 /* XXX I think this is bad news... -DaveM */
1095         hyper_clear_all_tags();
1096 #endif
1097
1098         put_ross_icr(HYPERSPARC_ICCR_FTD | HYPERSPARC_ICCR_ICE);
1099         hyper_flush_whole_icache();
1100         clear = srmmu_get_faddr();
1101         clear = srmmu_get_fstatus();
1102 }
1103
1104 static const struct sparc32_cachetlb_ops hypersparc_ops = {
1105         .cache_all      = hypersparc_flush_cache_all,
1106         .cache_mm       = hypersparc_flush_cache_mm,
1107         .cache_page     = hypersparc_flush_cache_page,
1108         .cache_range    = hypersparc_flush_cache_range,
1109         .tlb_all        = hypersparc_flush_tlb_all,
1110         .tlb_mm         = hypersparc_flush_tlb_mm,
1111         .tlb_page       = hypersparc_flush_tlb_page,
1112         .tlb_range      = hypersparc_flush_tlb_range,
1113         .page_to_ram    = hypersparc_flush_page_to_ram,
1114         .sig_insns      = hypersparc_flush_sig_insns,
1115         .page_for_dma   = hypersparc_flush_page_for_dma,
1116 };
1117
1118 static void __init init_hypersparc(void)
1119 {
1120         srmmu_name = "ROSS HyperSparc";
1121         srmmu_modtype = HyperSparc;
1122
1123         init_vac_layout();
1124
1125         is_hypersparc = 1;
1126         sparc32_cachetlb_ops = &hypersparc_ops;
1127
1128         poke_srmmu = poke_hypersparc;
1129
1130         hypersparc_setup_blockops();
1131 }
1132
1133 static void poke_swift(void)
1134 {
1135         unsigned long mreg;
1136
1137         /* Clear any crap from the cache or else... */
1138         swift_flush_cache_all();
1139
1140         /* Enable I & D caches */
1141         mreg = srmmu_get_mmureg();
1142         mreg |= (SWIFT_IE | SWIFT_DE);
1143         /*
1144          * The Swift branch folding logic is completely broken.  At
1145          * trap time, if things are just right, if can mistakenly
1146          * think that a trap is coming from kernel mode when in fact
1147          * it is coming from user mode (it mis-executes the branch in
1148          * the trap code).  So you see things like crashme completely
1149          * hosing your machine which is completely unacceptable.  Turn
1150          * this shit off... nice job Fujitsu.
1151          */
1152         mreg &= ~(SWIFT_BF);
1153         srmmu_set_mmureg(mreg);
1154 }
1155
1156 static const struct sparc32_cachetlb_ops swift_ops = {
1157         .cache_all      = swift_flush_cache_all,
1158         .cache_mm       = swift_flush_cache_mm,
1159         .cache_page     = swift_flush_cache_page,
1160         .cache_range    = swift_flush_cache_range,
1161         .tlb_all        = swift_flush_tlb_all,
1162         .tlb_mm         = swift_flush_tlb_mm,
1163         .tlb_page       = swift_flush_tlb_page,
1164         .tlb_range      = swift_flush_tlb_range,
1165         .page_to_ram    = swift_flush_page_to_ram,
1166         .sig_insns      = swift_flush_sig_insns,
1167         .page_for_dma   = swift_flush_page_for_dma,
1168 };
1169
1170 #define SWIFT_MASKID_ADDR  0x10003018
1171 static void __init init_swift(void)
1172 {
1173         unsigned long swift_rev;
1174
1175         __asm__ __volatile__("lda [%1] %2, %0\n\t"
1176                              "srl %0, 0x18, %0\n\t" :
1177                              "=r" (swift_rev) :
1178                              "r" (SWIFT_MASKID_ADDR), "i" (ASI_M_BYPASS));
1179         srmmu_name = "Fujitsu Swift";
1180         switch (swift_rev) {
1181         case 0x11:
1182         case 0x20:
1183         case 0x23:
1184         case 0x30:
1185                 srmmu_modtype = Swift_lots_o_bugs;
1186                 hwbug_bitmask |= (HWBUG_KERN_ACCBROKEN | HWBUG_KERN_CBITBROKEN);
1187                 /*
1188                  * Gee george, I wonder why Sun is so hush hush about
1189                  * this hardware bug... really braindamage stuff going
1190                  * on here.  However I think we can find a way to avoid
1191                  * all of the workaround overhead under Linux.  Basically,
1192                  * any page fault can cause kernel pages to become user
1193                  * accessible (the mmu gets confused and clears some of
1194                  * the ACC bits in kernel ptes).  Aha, sounds pretty
1195                  * horrible eh?  But wait, after extensive testing it appears
1196                  * that if you use pgd_t level large kernel pte's (like the
1197                  * 4MB pages on the Pentium) the bug does not get tripped
1198                  * at all.  This avoids almost all of the major overhead.
1199                  * Welcome to a world where your vendor tells you to,
1200                  * "apply this kernel patch" instead of "sorry for the
1201                  * broken hardware, send it back and we'll give you
1202                  * properly functioning parts"
1203                  */
1204                 break;
1205         case 0x25:
1206         case 0x31:
1207                 srmmu_modtype = Swift_bad_c;
1208                 hwbug_bitmask |= HWBUG_KERN_CBITBROKEN;
1209                 /*
1210                  * You see Sun allude to this hardware bug but never
1211                  * admit things directly, they'll say things like,
1212                  * "the Swift chip cache problems" or similar.
1213                  */
1214                 break;
1215         default:
1216                 srmmu_modtype = Swift_ok;
1217                 break;
1218         }
1219
1220         sparc32_cachetlb_ops = &swift_ops;
1221         flush_page_for_dma_global = 0;
1222
1223         /*
1224          * Are you now convinced that the Swift is one of the
1225          * biggest VLSI abortions of all time?  Bravo Fujitsu!
1226          * Fujitsu, the !#?!%$'d up processor people.  I bet if
1227          * you examined the microcode of the Swift you'd find
1228          * XXX's all over the place.
1229          */
1230         poke_srmmu = poke_swift;
1231 }
1232
1233 static void turbosparc_flush_cache_all(void)
1234 {
1235         flush_user_windows();
1236         turbosparc_idflash_clear();
1237 }
1238
1239 static void turbosparc_flush_cache_mm(struct mm_struct *mm)
1240 {
1241         FLUSH_BEGIN(mm)
1242         flush_user_windows();
1243         turbosparc_idflash_clear();
1244         FLUSH_END
1245 }
1246
1247 static void turbosparc_flush_cache_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
1248 {
1249         FLUSH_BEGIN(vma->vm_mm)
1250         flush_user_windows();
1251         turbosparc_idflash_clear();
1252         FLUSH_END
1253 }
1254
1255 static void turbosparc_flush_cache_page(struct vm_area_struct *vma, unsigned long page)
1256 {
1257         FLUSH_BEGIN(vma->vm_mm)
1258         flush_user_windows();
1259         if (vma->vm_flags & VM_EXEC)
1260                 turbosparc_flush_icache();
1261         turbosparc_flush_dcache();
1262         FLUSH_END
1263 }
1264
1265 /* TurboSparc is copy-back, if we turn it on, but this does not work. */
1266 static void turbosparc_flush_page_to_ram(unsigned long page)
1267 {
1268 #ifdef TURBOSPARC_WRITEBACK
1269         volatile unsigned long clear;
1270
1271         if (srmmu_probe(page))
1272                 turbosparc_flush_page_cache(page);
1273         clear = srmmu_get_fstatus();
1274 #endif
1275 }
1276
1277 static void turbosparc_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr)
1278 {
1279 }
1280
1281 static void turbosparc_flush_page_for_dma(unsigned long page)
1282 {
1283         turbosparc_flush_dcache();
1284 }
1285
1286 static void turbosparc_flush_tlb_all(void)
1287 {
1288         srmmu_flush_whole_tlb();
1289 }
1290
1291 static void turbosparc_flush_tlb_mm(struct mm_struct *mm)
1292 {
1293         FLUSH_BEGIN(mm)
1294         srmmu_flush_whole_tlb();
1295         FLUSH_END
1296 }
1297
1298 static void turbosparc_flush_tlb_range(struct vm_area_struct *vma, unsigned long start, unsigned long end)
1299 {
1300         FLUSH_BEGIN(vma->vm_mm)
1301         srmmu_flush_whole_tlb();
1302         FLUSH_END
1303 }
1304
1305 static void turbosparc_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
1306 {
1307         FLUSH_BEGIN(vma->vm_mm)
1308         srmmu_flush_whole_tlb();
1309         FLUSH_END
1310 }
1311
1312
1313 static void poke_turbosparc(void)
1314 {
1315         unsigned long mreg = srmmu_get_mmureg();
1316         unsigned long ccreg;
1317
1318         /* Clear any crap from the cache or else... */
1319         turbosparc_flush_cache_all();
1320         /* Temporarily disable I & D caches */
1321         mreg &= ~(TURBOSPARC_ICENABLE | TURBOSPARC_DCENABLE);
1322         mreg &= ~(TURBOSPARC_PCENABLE);         /* Don't check parity */
1323         srmmu_set_mmureg(mreg);
1324
1325         ccreg = turbosparc_get_ccreg();
1326
1327 #ifdef TURBOSPARC_WRITEBACK
1328         ccreg |= (TURBOSPARC_SNENABLE);         /* Do DVMA snooping in Dcache */
1329         ccreg &= ~(TURBOSPARC_uS2 | TURBOSPARC_WTENABLE);
1330                         /* Write-back D-cache, emulate VLSI
1331                          * abortion number three, not number one */
1332 #else
1333         /* For now let's play safe, optimize later */
1334         ccreg |= (TURBOSPARC_SNENABLE | TURBOSPARC_WTENABLE);
1335                         /* Do DVMA snooping in Dcache, Write-thru D-cache */
1336         ccreg &= ~(TURBOSPARC_uS2);
1337                         /* Emulate VLSI abortion number three, not number one */
1338 #endif
1339
1340         switch (ccreg & 7) {
1341         case 0: /* No SE cache */
1342         case 7: /* Test mode */
1343                 break;
1344         default:
1345                 ccreg |= (TURBOSPARC_SCENABLE);
1346         }
1347         turbosparc_set_ccreg(ccreg);
1348
1349         mreg |= (TURBOSPARC_ICENABLE | TURBOSPARC_DCENABLE); /* I & D caches on */
1350         mreg |= (TURBOSPARC_ICSNOOP);           /* Icache snooping on */
1351         srmmu_set_mmureg(mreg);
1352 }
1353
1354 static const struct sparc32_cachetlb_ops turbosparc_ops = {
1355         .cache_all      = turbosparc_flush_cache_all,
1356         .cache_mm       = turbosparc_flush_cache_mm,
1357         .cache_page     = turbosparc_flush_cache_page,
1358         .cache_range    = turbosparc_flush_cache_range,
1359         .tlb_all        = turbosparc_flush_tlb_all,
1360         .tlb_mm         = turbosparc_flush_tlb_mm,
1361         .tlb_page       = turbosparc_flush_tlb_page,
1362         .tlb_range      = turbosparc_flush_tlb_range,
1363         .page_to_ram    = turbosparc_flush_page_to_ram,
1364         .sig_insns      = turbosparc_flush_sig_insns,
1365         .page_for_dma   = turbosparc_flush_page_for_dma,
1366 };
1367
1368 static void __init init_turbosparc(void)
1369 {
1370         srmmu_name = "Fujitsu TurboSparc";
1371         srmmu_modtype = TurboSparc;
1372         sparc32_cachetlb_ops = &turbosparc_ops;
1373         poke_srmmu = poke_turbosparc;
1374 }
1375
1376 static void poke_tsunami(void)
1377 {
1378         unsigned long mreg = srmmu_get_mmureg();
1379
1380         tsunami_flush_icache();
1381         tsunami_flush_dcache();
1382         mreg &= ~TSUNAMI_ITD;
1383         mreg |= (TSUNAMI_IENAB | TSUNAMI_DENAB);
1384         srmmu_set_mmureg(mreg);
1385 }
1386
1387 static const struct sparc32_cachetlb_ops tsunami_ops = {
1388         .cache_all      = tsunami_flush_cache_all,
1389         .cache_mm       = tsunami_flush_cache_mm,
1390         .cache_page     = tsunami_flush_cache_page,
1391         .cache_range    = tsunami_flush_cache_range,
1392         .tlb_all        = tsunami_flush_tlb_all,
1393         .tlb_mm         = tsunami_flush_tlb_mm,
1394         .tlb_page       = tsunami_flush_tlb_page,
1395         .tlb_range      = tsunami_flush_tlb_range,
1396         .page_to_ram    = tsunami_flush_page_to_ram,
1397         .sig_insns      = tsunami_flush_sig_insns,
1398         .page_for_dma   = tsunami_flush_page_for_dma,
1399 };
1400
1401 static void __init init_tsunami(void)
1402 {
1403         /*
1404          * Tsunami's pretty sane, Sun and TI actually got it
1405          * somewhat right this time.  Fujitsu should have
1406          * taken some lessons from them.
1407          */
1408
1409         srmmu_name = "TI Tsunami";
1410         srmmu_modtype = Tsunami;
1411         sparc32_cachetlb_ops = &tsunami_ops;
1412         poke_srmmu = poke_tsunami;
1413
1414         tsunami_setup_blockops();
1415 }
1416
1417 static void poke_viking(void)
1418 {
1419         unsigned long mreg = srmmu_get_mmureg();
1420         static int smp_catch;
1421
1422         if (viking_mxcc_present) {
1423                 unsigned long mxcc_control = mxcc_get_creg();
1424
1425                 mxcc_control |= (MXCC_CTL_ECE | MXCC_CTL_PRE | MXCC_CTL_MCE);
1426                 mxcc_control &= ~(MXCC_CTL_RRC);
1427                 mxcc_set_creg(mxcc_control);
1428
1429                 /*
1430                  * We don't need memory parity checks.
1431                  * XXX This is a mess, have to dig out later. ecd.
1432                 viking_mxcc_turn_off_parity(&mreg, &mxcc_control);
1433                  */
1434
1435                 /* We do cache ptables on MXCC. */
1436                 mreg |= VIKING_TCENABLE;
1437         } else {
1438                 unsigned long bpreg;
1439
1440                 mreg &= ~(VIKING_TCENABLE);
1441                 if (smp_catch++) {
1442                         /* Must disable mixed-cmd mode here for other cpu's. */
1443                         bpreg = viking_get_bpreg();
1444                         bpreg &= ~(VIKING_ACTION_MIX);
1445                         viking_set_bpreg(bpreg);
1446
1447                         /* Just in case PROM does something funny. */
1448                         msi_set_sync();
1449                 }
1450         }
1451
1452         mreg |= VIKING_SPENABLE;
1453         mreg |= (VIKING_ICENABLE | VIKING_DCENABLE);
1454         mreg |= VIKING_SBENABLE;
1455         mreg &= ~(VIKING_ACENABLE);
1456         srmmu_set_mmureg(mreg);
1457 }
1458
1459 static struct sparc32_cachetlb_ops viking_ops __ro_after_init = {
1460         .cache_all      = viking_flush_cache_all,
1461         .cache_mm       = viking_flush_cache_mm,
1462         .cache_page     = viking_flush_cache_page,
1463         .cache_range    = viking_flush_cache_range,
1464         .tlb_all        = viking_flush_tlb_all,
1465         .tlb_mm         = viking_flush_tlb_mm,
1466         .tlb_page       = viking_flush_tlb_page,
1467         .tlb_range      = viking_flush_tlb_range,
1468         .page_to_ram    = viking_flush_page_to_ram,
1469         .sig_insns      = viking_flush_sig_insns,
1470         .page_for_dma   = viking_flush_page_for_dma,
1471 };
1472
1473 #ifdef CONFIG_SMP
1474 /* On sun4d the cpu broadcasts local TLB flushes, so we can just
1475  * perform the local TLB flush and all the other cpus will see it.
1476  * But, unfortunately, there is a bug in the sun4d XBUS backplane
1477  * that requires that we add some synchronization to these flushes.
1478  *
1479  * The bug is that the fifo which keeps track of all the pending TLB
1480  * broadcasts in the system is an entry or two too small, so if we
1481  * have too many going at once we'll overflow that fifo and lose a TLB
1482  * flush resulting in corruption.
1483  *
1484  * Our workaround is to take a global spinlock around the TLB flushes,
1485  * which guarentees we won't ever have too many pending.  It's a big
1486  * hammer, but a semaphore like system to make sure we only have N TLB
1487  * flushes going at once will require SMP locking anyways so there's
1488  * no real value in trying any harder than this.
1489  */
1490 static struct sparc32_cachetlb_ops viking_sun4d_smp_ops __ro_after_init = {
1491         .cache_all      = viking_flush_cache_all,
1492         .cache_mm       = viking_flush_cache_mm,
1493         .cache_page     = viking_flush_cache_page,
1494         .cache_range    = viking_flush_cache_range,
1495         .tlb_all        = sun4dsmp_flush_tlb_all,
1496         .tlb_mm         = sun4dsmp_flush_tlb_mm,
1497         .tlb_page       = sun4dsmp_flush_tlb_page,
1498         .tlb_range      = sun4dsmp_flush_tlb_range,
1499         .page_to_ram    = viking_flush_page_to_ram,
1500         .sig_insns      = viking_flush_sig_insns,
1501         .page_for_dma   = viking_flush_page_for_dma,
1502 };
1503 #endif
1504
1505 static void __init init_viking(void)
1506 {
1507         unsigned long mreg = srmmu_get_mmureg();
1508
1509         /* Ahhh, the viking.  SRMMU VLSI abortion number two... */
1510         if (mreg & VIKING_MMODE) {
1511                 srmmu_name = "TI Viking";
1512                 viking_mxcc_present = 0;
1513                 msi_set_sync();
1514
1515                 /*
1516                  * We need this to make sure old viking takes no hits
1517                  * on it's cache for dma snoops to workaround the
1518                  * "load from non-cacheable memory" interrupt bug.
1519                  * This is only necessary because of the new way in
1520                  * which we use the IOMMU.
1521                  */
1522                 viking_ops.page_for_dma = viking_flush_page;
1523 #ifdef CONFIG_SMP
1524                 viking_sun4d_smp_ops.page_for_dma = viking_flush_page;
1525 #endif
1526                 flush_page_for_dma_global = 0;
1527         } else {
1528                 srmmu_name = "TI Viking/MXCC";
1529                 viking_mxcc_present = 1;
1530                 srmmu_cache_pagetables = 1;
1531         }
1532
1533         sparc32_cachetlb_ops = (const struct sparc32_cachetlb_ops *)
1534                 &viking_ops;
1535 #ifdef CONFIG_SMP
1536         if (sparc_cpu_model == sun4d)
1537                 sparc32_cachetlb_ops = (const struct sparc32_cachetlb_ops *)
1538                         &viking_sun4d_smp_ops;
1539 #endif
1540
1541         poke_srmmu = poke_viking;
1542 }
1543
1544 /* Probe for the srmmu chip version. */
1545 static void __init get_srmmu_type(void)
1546 {
1547         unsigned long mreg, psr;
1548         unsigned long mod_typ, mod_rev, psr_typ, psr_vers;
1549
1550         srmmu_modtype = SRMMU_INVAL_MOD;
1551         hwbug_bitmask = 0;
1552
1553         mreg = srmmu_get_mmureg(); psr = get_psr();
1554         mod_typ = (mreg & 0xf0000000) >> 28;
1555         mod_rev = (mreg & 0x0f000000) >> 24;
1556         psr_typ = (psr >> 28) & 0xf;
1557         psr_vers = (psr >> 24) & 0xf;
1558
1559         /* First, check for sparc-leon. */
1560         if (sparc_cpu_model == sparc_leon) {
1561                 init_leon();
1562                 return;
1563         }
1564
1565         /* Second, check for HyperSparc or Cypress. */
1566         if (mod_typ == 1) {
1567                 switch (mod_rev) {
1568                 case 7:
1569                         /* UP or MP Hypersparc */
1570                         init_hypersparc();
1571                         break;
1572                 case 0:
1573                 case 2:
1574                 case 10:
1575                 case 11:
1576                 case 12:
1577                 case 13:
1578                 case 14:
1579                 case 15:
1580                 default:
1581                         prom_printf("Sparc-Linux Cypress support does not longer exit.\n");
1582                         prom_halt();
1583                         break;
1584                 }
1585                 return;
1586         }
1587
1588         /* Now Fujitsu TurboSparc. It might happen that it is
1589          * in Swift emulation mode, so we will check later...
1590          */
1591         if (psr_typ == 0 && psr_vers == 5) {
1592                 init_turbosparc();
1593                 return;
1594         }
1595
1596         /* Next check for Fujitsu Swift. */
1597         if (psr_typ == 0 && psr_vers == 4) {
1598                 phandle cpunode;
1599                 char node_str[128];
1600
1601                 /* Look if it is not a TurboSparc emulating Swift... */
1602                 cpunode = prom_getchild(prom_root_node);
1603                 while ((cpunode = prom_getsibling(cpunode)) != 0) {
1604                         prom_getstring(cpunode, "device_type", node_str, sizeof(node_str));
1605                         if (!strcmp(node_str, "cpu")) {
1606                                 if (!prom_getintdefault(cpunode, "psr-implementation", 1) &&
1607                                     prom_getintdefault(cpunode, "psr-version", 1) == 5) {
1608                                         init_turbosparc();
1609                                         return;
1610                                 }
1611                                 break;
1612                         }
1613                 }
1614
1615                 init_swift();
1616                 return;
1617         }
1618
1619         /* Now the Viking family of srmmu. */
1620         if (psr_typ == 4 &&
1621            ((psr_vers == 0) ||
1622             ((psr_vers == 1) && (mod_typ == 0) && (mod_rev == 0)))) {
1623                 init_viking();
1624                 return;
1625         }
1626
1627         /* Finally the Tsunami. */
1628         if (psr_typ == 4 && psr_vers == 1 && (mod_typ || mod_rev)) {
1629                 init_tsunami();
1630                 return;
1631         }
1632
1633         /* Oh well */
1634         srmmu_is_bad();
1635 }
1636
1637 #ifdef CONFIG_SMP
1638 /* Local cross-calls. */
1639 static void smp_flush_page_for_dma(unsigned long page)
1640 {
1641         xc1((smpfunc_t) local_ops->page_for_dma, page);
1642         local_ops->page_for_dma(page);
1643 }
1644
1645 static void smp_flush_cache_all(void)
1646 {
1647         xc0((smpfunc_t) local_ops->cache_all);
1648         local_ops->cache_all();
1649 }
1650
1651 static void smp_flush_tlb_all(void)
1652 {
1653         xc0((smpfunc_t) local_ops->tlb_all);
1654         local_ops->tlb_all();
1655 }
1656
1657 static void smp_flush_cache_mm(struct mm_struct *mm)
1658 {
1659         if (mm->context != NO_CONTEXT) {
1660                 cpumask_t cpu_mask;
1661                 cpumask_copy(&cpu_mask, mm_cpumask(mm));
1662                 cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1663                 if (!cpumask_empty(&cpu_mask))
1664                         xc1((smpfunc_t) local_ops->cache_mm, (unsigned long) mm);
1665                 local_ops->cache_mm(mm);
1666         }
1667 }
1668
1669 static void smp_flush_tlb_mm(struct mm_struct *mm)
1670 {
1671         if (mm->context != NO_CONTEXT) {
1672                 cpumask_t cpu_mask;
1673                 cpumask_copy(&cpu_mask, mm_cpumask(mm));
1674                 cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1675                 if (!cpumask_empty(&cpu_mask)) {
1676                         xc1((smpfunc_t) local_ops->tlb_mm, (unsigned long) mm);
1677                         if (atomic_read(&mm->mm_users) == 1 && current->active_mm == mm)
1678                                 cpumask_copy(mm_cpumask(mm),
1679                                              cpumask_of(smp_processor_id()));
1680                 }
1681                 local_ops->tlb_mm(mm);
1682         }
1683 }
1684
1685 static void smp_flush_cache_range(struct vm_area_struct *vma,
1686                                   unsigned long start,
1687                                   unsigned long end)
1688 {
1689         struct mm_struct *mm = vma->vm_mm;
1690
1691         if (mm->context != NO_CONTEXT) {
1692                 cpumask_t cpu_mask;
1693                 cpumask_copy(&cpu_mask, mm_cpumask(mm));
1694                 cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1695                 if (!cpumask_empty(&cpu_mask))
1696                         xc3((smpfunc_t) local_ops->cache_range,
1697                             (unsigned long) vma, start, end);
1698                 local_ops->cache_range(vma, start, end);
1699         }
1700 }
1701
1702 static void smp_flush_tlb_range(struct vm_area_struct *vma,
1703                                 unsigned long start,
1704                                 unsigned long end)
1705 {
1706         struct mm_struct *mm = vma->vm_mm;
1707
1708         if (mm->context != NO_CONTEXT) {
1709                 cpumask_t cpu_mask;
1710                 cpumask_copy(&cpu_mask, mm_cpumask(mm));
1711                 cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1712                 if (!cpumask_empty(&cpu_mask))
1713                         xc3((smpfunc_t) local_ops->tlb_range,
1714                             (unsigned long) vma, start, end);
1715                 local_ops->tlb_range(vma, start, end);
1716         }
1717 }
1718
1719 static void smp_flush_cache_page(struct vm_area_struct *vma, unsigned long page)
1720 {
1721         struct mm_struct *mm = vma->vm_mm;
1722
1723         if (mm->context != NO_CONTEXT) {
1724                 cpumask_t cpu_mask;
1725                 cpumask_copy(&cpu_mask, mm_cpumask(mm));
1726                 cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1727                 if (!cpumask_empty(&cpu_mask))
1728                         xc2((smpfunc_t) local_ops->cache_page,
1729                             (unsigned long) vma, page);
1730                 local_ops->cache_page(vma, page);
1731         }
1732 }
1733
1734 static void smp_flush_tlb_page(struct vm_area_struct *vma, unsigned long page)
1735 {
1736         struct mm_struct *mm = vma->vm_mm;
1737
1738         if (mm->context != NO_CONTEXT) {
1739                 cpumask_t cpu_mask;
1740                 cpumask_copy(&cpu_mask, mm_cpumask(mm));
1741                 cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1742                 if (!cpumask_empty(&cpu_mask))
1743                         xc2((smpfunc_t) local_ops->tlb_page,
1744                             (unsigned long) vma, page);
1745                 local_ops->tlb_page(vma, page);
1746         }
1747 }
1748
1749 static void smp_flush_page_to_ram(unsigned long page)
1750 {
1751         /* Current theory is that those who call this are the one's
1752          * who have just dirtied their cache with the pages contents
1753          * in kernel space, therefore we only run this on local cpu.
1754          *
1755          * XXX This experiment failed, research further... -DaveM
1756          */
1757 #if 1
1758         xc1((smpfunc_t) local_ops->page_to_ram, page);
1759 #endif
1760         local_ops->page_to_ram(page);
1761 }
1762
1763 static void smp_flush_sig_insns(struct mm_struct *mm, unsigned long insn_addr)
1764 {
1765         cpumask_t cpu_mask;
1766         cpumask_copy(&cpu_mask, mm_cpumask(mm));
1767         cpumask_clear_cpu(smp_processor_id(), &cpu_mask);
1768         if (!cpumask_empty(&cpu_mask))
1769                 xc2((smpfunc_t) local_ops->sig_insns,
1770                     (unsigned long) mm, insn_addr);
1771         local_ops->sig_insns(mm, insn_addr);
1772 }
1773
1774 static struct sparc32_cachetlb_ops smp_cachetlb_ops __ro_after_init = {
1775         .cache_all      = smp_flush_cache_all,
1776         .cache_mm       = smp_flush_cache_mm,
1777         .cache_page     = smp_flush_cache_page,
1778         .cache_range    = smp_flush_cache_range,
1779         .tlb_all        = smp_flush_tlb_all,
1780         .tlb_mm         = smp_flush_tlb_mm,
1781         .tlb_page       = smp_flush_tlb_page,
1782         .tlb_range      = smp_flush_tlb_range,
1783         .page_to_ram    = smp_flush_page_to_ram,
1784         .sig_insns      = smp_flush_sig_insns,
1785         .page_for_dma   = smp_flush_page_for_dma,
1786 };
1787 #endif
1788
1789 /* Load up routines and constants for sun4m and sun4d mmu */
1790 void __init load_mmu(void)
1791 {
1792         /* Functions */
1793         get_srmmu_type();
1794
1795 #ifdef CONFIG_SMP
1796         /* El switcheroo... */
1797         local_ops = sparc32_cachetlb_ops;
1798
1799         if (sparc_cpu_model == sun4d || sparc_cpu_model == sparc_leon) {
1800                 smp_cachetlb_ops.tlb_all = local_ops->tlb_all;
1801                 smp_cachetlb_ops.tlb_mm = local_ops->tlb_mm;
1802                 smp_cachetlb_ops.tlb_range = local_ops->tlb_range;
1803                 smp_cachetlb_ops.tlb_page = local_ops->tlb_page;
1804         }
1805
1806         if (poke_srmmu == poke_viking) {
1807                 /* Avoid unnecessary cross calls. */
1808                 smp_cachetlb_ops.cache_all = local_ops->cache_all;
1809                 smp_cachetlb_ops.cache_mm = local_ops->cache_mm;
1810                 smp_cachetlb_ops.cache_range = local_ops->cache_range;
1811                 smp_cachetlb_ops.cache_page = local_ops->cache_page;
1812
1813                 smp_cachetlb_ops.page_to_ram = local_ops->page_to_ram;
1814                 smp_cachetlb_ops.sig_insns = local_ops->sig_insns;
1815                 smp_cachetlb_ops.page_for_dma = local_ops->page_for_dma;
1816         }
1817
1818         /* It really is const after this point. */
1819         sparc32_cachetlb_ops = (const struct sparc32_cachetlb_ops *)
1820                 &smp_cachetlb_ops;
1821 #endif
1822
1823         if (sparc_cpu_model != sun4d)
1824                 ld_mmu_iommu();
1825 #ifdef CONFIG_SMP
1826         if (sparc_cpu_model == sun4d)
1827                 sun4d_init_smp();
1828         else if (sparc_cpu_model == sparc_leon)
1829                 leon_init_smp();
1830         else
1831                 sun4m_init_smp();
1832 #endif
1833 }