arch/sparc64/mm/init.c

   1 /*  $Id: init.c,v 1.209 2002/02/09 19:49:31 davem Exp $
   2  *  arch/sparc64/mm/init.c
   3  *
   4  *  Copyright (C) 1996-1999 David S. Miller (davem@caip.rutgers.edu)
   5  *  Copyright (C) 1997-1999 Jakub Jelinek (jj@sunsite.mff.cuni.cz)
   6  */
   7
   8 #include <linux/module.h>
   9 #include <linux/kernel.h>
  10 #include <linux/sched.h>
  11 #include <linux/string.h>
  12 #include <linux/init.h>
  13 #include <linux/bootmem.h>
  14 #include <linux/mm.h>
  15 #include <linux/hugetlb.h>
  16 #include <linux/slab.h>
  17 #include <linux/initrd.h>
  18 #include <linux/swap.h>
  19 #include <linux/pagemap.h>
  20 #include <linux/poison.h>
  21 #include <linux/fs.h>
  22 #include <linux/seq_file.h>
  23 #include <linux/kprobes.h>
  24 #include <linux/cache.h>
  25 #include <linux/sort.h>
  26 #include <linux/percpu.h>
  27 #include <linux/lmb.h>
  28 #include <linux/mmzone.h>
  29
  30 #include <asm/head.h>
  31 #include <asm/system.h>
  32 #include <asm/page.h>
  33 #include <asm/pgalloc.h>
  34 #include <asm/pgtable.h>
  35 #include <asm/oplib.h>
  36 #include <asm/iommu.h>
  37 #include <asm/io.h>
  38 #include <asm/uaccess.h>
  39 #include <asm/mmu_context.h>
  40 #include <asm/tlbflush.h>
  41 #include <asm/dma.h>
  42 #include <asm/starfire.h>
  43 #include <asm/tlb.h>
  44 #include <asm/spitfire.h>
  45 #include <asm/sections.h>
  46 #include <asm/tsb.h>
  47 #include <asm/hypervisor.h>
  48 #include <asm/prom.h>
  49 #include <asm/sstate.h>
  50 #include <asm/mdesc.h>
  51 #include <asm/cpudata.h>
  52
  53 #define MAX_PHYS_ADDRESS        (1UL << 42UL)
  54 #define KPTE_BITMAP_CHUNK_SZ    (256UL * 1024UL * 1024UL)
  55 #define KPTE_BITMAP_BYTES       \
  56         ((MAX_PHYS_ADDRESS / KPTE_BITMAP_CHUNK_SZ) / 8)
  57
  58 unsigned long kern_linear_pte_xor[2] __read_mostly;
  59
  60 /* A bitmap, one bit for every 256MB of physical memory.  If the bit
  61  * is clear, we should use a 4MB page (via kern_linear_pte_xor[0]) else
  62  * if set we should use a 256MB page (via kern_linear_pte_xor[1]).
  63  */
  64 unsigned long kpte_linear_bitmap[KPTE_BITMAP_BYTES / sizeof(unsigned long)];
  65
  66 #ifndef CONFIG_DEBUG_PAGEALLOC
  67 /* A special kernel TSB for 4MB and 256MB linear mappings.
  68  * Space is allocated for this right after the trap table
  69  * in arch/sparc64/kernel/head.S
  70  */
  71 extern struct tsb swapper_4m_tsb[KERNEL_TSB4M_NENTRIES];
  72 #endif
  73
  74 #define MAX_BANKS       32
  75
  76 static struct linux_prom64_registers pavail[MAX_BANKS] __initdata;
  77 static int pavail_ents __initdata;
  78
  79 static int cmp_p64(const void *a, const void *b)
  80 {
  81         const struct linux_prom64_registers *x = a, *y = b;
  82
  83         if (x->phys_addr > y->phys_addr)
  84                 return 1;
  85         if (x->phys_addr < y->phys_addr)
  86                 return -1;
  87         return 0;
  88 }
  89
  90 static void __init read_obp_memory(const char *property,
  91                                    struct linux_prom64_registers *regs,
  92                                    int *num_ents)
  93 {
  94         int node = prom_finddevice("/memory");
  95         int prop_size = prom_getproplen(node, property);
  96         int ents, ret, i;
  97
  98         ents = prop_size / sizeof(struct linux_prom64_registers);
  99         if (ents > MAX_BANKS) {
 100                 prom_printf("The machine has more %s property entries than "
 101                             "this kernel can support (%d).\n",
 102                             property, MAX_BANKS);
 103                 prom_halt();
 104         }
 105
 106         ret = prom_getproperty(node, property, (char *) regs, prop_size);
 107         if (ret == -1) {
 108                 prom_printf("Couldn't get %s property from /memory.\n");
 109                 prom_halt();
 110         }
 111
 112         /* Sanitize what we got from the firmware, by page aligning
 113          * everything.
 114          */
 115         for (i = 0; i < ents; i++) {
 116                 unsigned long base, size;
 117
 118                 base = regs[i].phys_addr;
 119                 size = regs[i].reg_size;
 120
 121                 size &= PAGE_MASK;
 122                 if (base & ~PAGE_MASK) {
 123                         unsigned long new_base = PAGE_ALIGN(base);
 124
 125                         size -= new_base - base;
 126                         if ((long) size < 0L)
 127                                 size = 0UL;
 128                         base = new_base;
 129                 }
 130                 if (size == 0UL) {
 131                         /* If it is empty, simply get rid of it.
 132                          * This simplifies the logic of the other
 133                          * functions that process these arrays.
 134                          */
 135                         memmove(&regs[i], &regs[i + 1],
 136                                 (ents - i - 1) * sizeof(regs[0]));
 137                         i--;
 138                         ents--;
 139                         continue;
 140                 }
 141                 regs[i].phys_addr = base;
 142                 regs[i].reg_size = size;
 143         }
 144
 145         *num_ents = ents;
 146
 147         sort(regs, ents, sizeof(struct linux_prom64_registers),
 148              cmp_p64, NULL);
 149 }
 150
 151 unsigned long *sparc64_valid_addr_bitmap __read_mostly;
 152
 153 /* Kernel physical address base and size in bytes.  */
 154 unsigned long kern_base __read_mostly;
 155 unsigned long kern_size __read_mostly;
 156
 157 /* Initial ramdisk setup */
 158 extern unsigned long sparc_ramdisk_image64;
 159 extern unsigned int sparc_ramdisk_image;
 160 extern unsigned int sparc_ramdisk_size;
 161
 162 struct page *mem_map_zero __read_mostly;
 163
 164 unsigned int sparc64_highest_unlocked_tlb_ent __read_mostly;
 165
 166 unsigned long sparc64_kern_pri_context __read_mostly;
 167 unsigned long sparc64_kern_pri_nuc_bits __read_mostly;
 168 unsigned long sparc64_kern_sec_context __read_mostly;
 169
 170 int num_kernel_image_mappings;
 171
 172 #ifdef CONFIG_DEBUG_DCFLUSH
 173 atomic_t dcpage_flushes = ATOMIC_INIT(0);
 174 #ifdef CONFIG_SMP
 175 atomic_t dcpage_flushes_xcall = ATOMIC_INIT(0);
 176 #endif
 177 #endif
 178
 179 inline void flush_dcache_page_impl(struct page *page)
 180 {
 181         BUG_ON(tlb_type == hypervisor);
 182 #ifdef CONFIG_DEBUG_DCFLUSH
 183         atomic_inc(&dcpage_flushes);
 184 #endif
 185
 186 #ifdef DCACHE_ALIASING_POSSIBLE
 187         __flush_dcache_page(page_address(page),
 188                             ((tlb_type == spitfire) &&
 189                              page_mapping(page) != NULL));
 190 #else
 191         if (page_mapping(page) != NULL &&
 192             tlb_type == spitfire)
 193                 __flush_icache_page(__pa(page_address(page)));
 194 #endif
 195 }
 196
 197 #define PG_dcache_dirty         PG_arch_1
 198 #define PG_dcache_cpu_shift     32UL
 199 #define PG_dcache_cpu_mask      \
 200         ((1UL<<ilog2(roundup_pow_of_two(NR_CPUS)))-1UL)
 201
 202 #define dcache_dirty_cpu(page) \
 203         (((page)->flags >> PG_dcache_cpu_shift) & PG_dcache_cpu_mask)
 204
 205 static inline void set_dcache_dirty(struct page *page, int this_cpu)
 206 {
 207         unsigned long mask = this_cpu;
 208         unsigned long non_cpu_bits;
 209
 210         non_cpu_bits = ~(PG_dcache_cpu_mask << PG_dcache_cpu_shift);
 211         mask = (mask << PG_dcache_cpu_shift) | (1UL << PG_dcache_dirty);
 212
 213         __asm__ __volatile__("1:\n\t"
 214                              "ldx       [%2], %%g7\n\t"
 215                              "and       %%g7, %1, %%g1\n\t"
 216                              "or        %%g1, %0, %%g1\n\t"
 217                              "casx      [%2], %%g7, %%g1\n\t"
 218                              "cmp       %%g7, %%g1\n\t"
 219                              "membar    #StoreLoad | #StoreStore\n\t"
 220                              "bne,pn    %%xcc, 1b\n\t"
 221                              " nop"
 222                              : /* no outputs */
 223                              : "r" (mask), "r" (non_cpu_bits), "r" (&page->flags)
 224                              : "g1", "g7");
 225 }
 226
 227 static inline void clear_dcache_dirty_cpu(struct page *page, unsigned long cpu)
 228 {
 229         unsigned long mask = (1UL << PG_dcache_dirty);
 230
 231         __asm__ __volatile__("! test_and_clear_dcache_dirty\n"
 232                              "1:\n\t"
 233                              "ldx       [%2], %%g7\n\t"
 234                              "srlx      %%g7, %4, %%g1\n\t"
 235                              "and       %%g1, %3, %%g1\n\t"
 236                              "cmp       %%g1, %0\n\t"
 237                              "bne,pn    %%icc, 2f\n\t"
 238                              " andn     %%g7, %1, %%g1\n\t"
 239                              "casx      [%2], %%g7, %%g1\n\t"
 240                              "cmp       %%g7, %%g1\n\t"
 241                              "membar    #StoreLoad | #StoreStore\n\t"
 242                              "bne,pn    %%xcc, 1b\n\t"
 243                              " nop\n"
 244                              "2:"
 245                              : /* no outputs */
 246                              : "r" (cpu), "r" (mask), "r" (&page->flags),
 247                                "i" (PG_dcache_cpu_mask),
 248                                "i" (PG_dcache_cpu_shift)
 249                              : "g1", "g7");
 250 }
 251
 252 static inline void tsb_insert(struct tsb *ent, unsigned long tag, unsigned long pte)
 253 {
 254         unsigned long tsb_addr = (unsigned long) ent;
 255
 256         if (tlb_type == cheetah_plus || tlb_type == hypervisor)
 257                 tsb_addr = __pa(tsb_addr);
 258
 259         __tsb_insert(tsb_addr, tag, pte);
 260 }
 261
 262 unsigned long _PAGE_ALL_SZ_BITS __read_mostly;
 263 unsigned long _PAGE_SZBITS __read_mostly;
 264
 265 void update_mmu_cache(struct vm_area_struct *vma, unsigned long address, pte_t pte)
 266 {
 267         struct mm_struct *mm;
 268         struct tsb *tsb;
 269         unsigned long tag, flags;
 270         unsigned long tsb_index, tsb_hash_shift;
 271
 272         if (tlb_type != hypervisor) {
 273                 unsigned long pfn = pte_pfn(pte);
 274                 unsigned long pg_flags;
 275                 struct page *page;
 276
 277                 if (pfn_valid(pfn) &&
 278                     (page = pfn_to_page(pfn), page_mapping(page)) &&
 279                     ((pg_flags = page->flags) & (1UL << PG_dcache_dirty))) {
 280                         int cpu = ((pg_flags >> PG_dcache_cpu_shift) &
 281                                    PG_dcache_cpu_mask);
 282                         int this_cpu = get_cpu();
 283
 284                         /* This is just to optimize away some function calls
 285                          * in the SMP case.
 286                          */
 287                         if (cpu == this_cpu)
 288                                 flush_dcache_page_impl(page);
 289                         else
 290                                 smp_flush_dcache_page_impl(page, cpu);
 291
 292                         clear_dcache_dirty_cpu(page, cpu);
 293
 294                         put_cpu();
 295                 }
 296         }
 297
 298         mm = vma->vm_mm;
 299
 300         tsb_index = MM_TSB_BASE;
 301         tsb_hash_shift = PAGE_SHIFT;
 302
 303         spin_lock_irqsave(&mm->context.lock, flags);
 304
 305 #ifdef CONFIG_HUGETLB_PAGE
 306         if (mm->context.tsb_block[MM_TSB_HUGE].tsb != NULL) {
 307                 if ((tlb_type == hypervisor &&
 308                      (pte_val(pte) & _PAGE_SZALL_4V) == _PAGE_SZHUGE_4V) ||
 309                     (tlb_type != hypervisor &&
 310                      (pte_val(pte) & _PAGE_SZALL_4U) == _PAGE_SZHUGE_4U)) {
 311                         tsb_index = MM_TSB_HUGE;
 312                         tsb_hash_shift = HPAGE_SHIFT;
 313                 }
 314         }
 315 #endif
 316
 317         tsb = mm->context.tsb_block[tsb_index].tsb;
 318         tsb += ((address >> tsb_hash_shift) &
 319                 (mm->context.tsb_block[tsb_index].tsb_nentries - 1UL));
 320         tag = (address >> 22UL);
 321         tsb_insert(tsb, tag, pte_val(pte));
 322
 323         spin_unlock_irqrestore(&mm->context.lock, flags);
 324 }
 325
 326 void flush_dcache_page(struct page *page)
 327 {
 328         struct address_space *mapping;
 329         int this_cpu;
 330
 331         if (tlb_type == hypervisor)
 332                 return;
 333
 334         /* Do not bother with the expensive D-cache flush if it
 335          * is merely the zero page.  The 'bigcore' testcase in GDB
 336          * causes this case to run millions of times.
 337          */
 338         if (page == ZERO_PAGE(0))
 339                 return;
 340
 341         this_cpu = get_cpu();
 342
 343         mapping = page_mapping(page);
 344         if (mapping && !mapping_mapped(mapping)) {
 345                 int dirty = test_bit(PG_dcache_dirty, &page->flags);
 346                 if (dirty) {
 347                         int dirty_cpu = dcache_dirty_cpu(page);
 348
 349                         if (dirty_cpu == this_cpu)
 350                                 goto out;
 351                         smp_flush_dcache_page_impl(page, dirty_cpu);
 352                 }
 353                 set_dcache_dirty(page, this_cpu);
 354         } else {
 355                 /* We could delay the flush for the !page_mapping
 356                  * case too.  But that case is for exec env/arg
 357                  * pages and those are %99 certainly going to get
 358                  * faulted into the tlb (and thus flushed) anyways.
 359                  */
 360                 flush_dcache_page_impl(page);
 361         }
 362
 363 out:
 364         put_cpu();
 365 }
 366
 367 void __kprobes flush_icache_range(unsigned long start, unsigned long end)
 368 {
 369         /* Cheetah and Hypervisor platform cpus have coherent I-cache. */
 370         if (tlb_type == spitfire) {
 371                 unsigned long kaddr;
 372
 373                 /* This code only runs on Spitfire cpus so this is
 374                  * why we can assume _PAGE_PADDR_4U.
 375                  */
 376                 for (kaddr = start; kaddr < end; kaddr += PAGE_SIZE) {
 377                         unsigned long paddr, mask = _PAGE_PADDR_4U;
 378
 379                         if (kaddr >= PAGE_OFFSET)
 380                                 paddr = kaddr & mask;
 381                         else {
 382                                 pgd_t *pgdp = pgd_offset_k(kaddr);
 383                                 pud_t *pudp = pud_offset(pgdp, kaddr);
 384                                 pmd_t *pmdp = pmd_offset(pudp, kaddr);
 385                                 pte_t *ptep = pte_offset_kernel(pmdp, kaddr);
 386
 387                                 paddr = pte_val(*ptep) & mask;
 388                         }
 389                         __flush_icache_page(paddr);
 390                 }
 391         }
 392 }
 393
 394 void show_mem(void)
 395 {
 396         unsigned long total = 0, reserved = 0;
 397         unsigned long shared = 0, cached = 0;
 398         pg_data_t *pgdat;
 399
 400         printk(KERN_INFO "Mem-info:\n");
 401         show_free_areas();
 402         printk(KERN_INFO "Free swap:       %6ldkB\n",
 403                nr_swap_pages << (PAGE_SHIFT-10));
 404         for_each_online_pgdat(pgdat) {
 405                 unsigned long i, flags;
 406
 407                 pgdat_resize_lock(pgdat, &flags);
 408                 for (i = 0; i < pgdat->node_spanned_pages; i++) {
 409                         struct page *page = pgdat_page_nr(pgdat, i);
 410                         total++;
 411                         if (PageReserved(page))
 412                                 reserved++;
 413                         else if (PageSwapCache(page))
 414                                 cached++;
 415                         else if (page_count(page))
 416                                 shared += page_count(page) - 1;
 417                 }
 418                 pgdat_resize_unlock(pgdat, &flags);
 419         }
 420
 421         printk(KERN_INFO "%lu pages of RAM\n", total);
 422         printk(KERN_INFO "%lu reserved pages\n", reserved);
 423         printk(KERN_INFO "%lu pages shared\n", shared);
 424         printk(KERN_INFO "%lu pages swap cached\n", cached);
 425
 426         printk(KERN_INFO "%lu pages dirty\n",
 427                global_page_state(NR_FILE_DIRTY));
 428         printk(KERN_INFO "%lu pages writeback\n",
 429                global_page_state(NR_WRITEBACK));
 430         printk(KERN_INFO "%lu pages mapped\n",
 431                global_page_state(NR_FILE_MAPPED));
 432         printk(KERN_INFO "%lu pages slab\n",
 433                 global_page_state(NR_SLAB_RECLAIMABLE) +
 434                 global_page_state(NR_SLAB_UNRECLAIMABLE));
 435         printk(KERN_INFO "%lu pages pagetables\n",
 436                global_page_state(NR_PAGETABLE));
 437 }
 438
 439 void mmu_info(struct seq_file *m)
 440 {
 441         if (tlb_type == cheetah)
 442                 seq_printf(m, "MMU Type\t: Cheetah\n");
 443         else if (tlb_type == cheetah_plus)
 444                 seq_printf(m, "MMU Type\t: Cheetah+\n");
 445         else if (tlb_type == spitfire)
 446                 seq_printf(m, "MMU Type\t: Spitfire\n");
 447         else if (tlb_type == hypervisor)
 448                 seq_printf(m, "MMU Type\t: Hypervisor (sun4v)\n");
 449         else
 450                 seq_printf(m, "MMU Type\t: ???\n");
 451
 452 #ifdef CONFIG_DEBUG_DCFLUSH
 453         seq_printf(m, "DCPageFlushes\t: %d\n",
 454                    atomic_read(&dcpage_flushes));
 455 #ifdef CONFIG_SMP
 456         seq_printf(m, "DCPageFlushesXC\t: %d\n",
 457                    atomic_read(&dcpage_flushes_xcall));
 458 #endif /* CONFIG_SMP */
 459 #endif /* CONFIG_DEBUG_DCFLUSH */
 460 }
 461
 462 struct linux_prom_translation {
 463         unsigned long virt;
 464         unsigned long size;
 465         unsigned long data;
 466 };
 467
 468 /* Exported for kernel TLB miss handling in ktlb.S */
 469 struct linux_prom_translation prom_trans[512] __read_mostly;
 470 unsigned int prom_trans_ents __read_mostly;
 471
 472 /* Exported for SMP bootup purposes. */
 473 unsigned long kern_locked_tte_data;
 474
 475 /* The obp translations are saved based on 8k pagesize, since obp can
 476  * use a mixture of pagesizes. Misses to the LOW_OBP_ADDRESS ->
 477  * HI_OBP_ADDRESS range are handled in ktlb.S.
 478  */
 479 static inline int in_obp_range(unsigned long vaddr)
 480 {
 481         return (vaddr >= LOW_OBP_ADDRESS &&
 482                 vaddr < HI_OBP_ADDRESS);
 483 }
 484
 485 static int cmp_ptrans(const void *a, const void *b)
 486 {
 487         const struct linux_prom_translation *x = a, *y = b;
 488
 489         if (x->virt > y->virt)
 490                 return 1;
 491         if (x->virt < y->virt)
 492                 return -1;
 493         return 0;
 494 }
 495
 496 /* Read OBP translations property into 'prom_trans[]'.  */
 497 static void __init read_obp_translations(void)
 498 {
 499         int n, node, ents, first, last, i;
 500
 501         node = prom_finddevice("/virtual-memory");
 502         n = prom_getproplen(node, "translations");
 503         if (unlikely(n == 0 || n == -1)) {
 504                 prom_printf("prom_mappings: Couldn't get size.\n");
 505                 prom_halt();
 506         }
 507         if (unlikely(n > sizeof(prom_trans))) {
 508                 prom_printf("prom_mappings: Size %Zd is too big.\n", n);
 509                 prom_halt();
 510         }
 511
 512         if ((n = prom_getproperty(node, "translations",
 513                                   (char *)&prom_trans[0],
 514                                   sizeof(prom_trans))) == -1) {
 515                 prom_printf("prom_mappings: Couldn't get property.\n");
 516                 prom_halt();
 517         }
 518
 519         n = n / sizeof(struct linux_prom_translation);
 520
 521         ents = n;
 522
 523         sort(prom_trans, ents, sizeof(struct linux_prom_translation),
 524              cmp_ptrans, NULL);
 525
 526         /* Now kick out all the non-OBP entries.  */
 527         for (i = 0; i < ents; i++) {
 528                 if (in_obp_range(prom_trans[i].virt))
 529                         break;
 530         }
 531         first = i;
 532         for (; i < ents; i++) {
 533                 if (!in_obp_range(prom_trans[i].virt))
 534                         break;
 535         }
 536         last = i;
 537
 538         for (i = 0; i < (last - first); i++) {
 539                 struct linux_prom_translation *src = &prom_trans[i + first];
 540                 struct linux_prom_translation *dest = &prom_trans[i];
 541
 542                 *dest = *src;
 543         }
 544         for (; i < ents; i++) {
 545                 struct linux_prom_translation *dest = &prom_trans[i];
 546                 dest->virt = dest->size = dest->data = 0x0UL;
 547         }
 548
 549         prom_trans_ents = last - first;
 550
 551         if (tlb_type == spitfire) {
 552                 /* Clear diag TTE bits. */
 553                 for (i = 0; i < prom_trans_ents; i++)
 554                         prom_trans[i].data &= ~0x0003fe0000000000UL;
 555         }
 556 }
 557
 558 static void __init hypervisor_tlb_lock(unsigned long vaddr,
 559                                        unsigned long pte,
 560                                        unsigned long mmu)
 561 {
 562         unsigned long ret = sun4v_mmu_map_perm_addr(vaddr, 0, pte, mmu);
 563
 564         if (ret != 0) {
 565                 prom_printf("hypervisor_tlb_lock[%lx:%lx:%lx:%lx]: "
 566                             "errors with %lx\n", vaddr, 0, pte, mmu, ret);
 567                 prom_halt();
 568         }
 569 }
 570
 571 static unsigned long kern_large_tte(unsigned long paddr);
 572
 573 static void __init remap_kernel(void)
 574 {
 575         unsigned long phys_page, tte_vaddr, tte_data;
 576         int i, tlb_ent = sparc64_highest_locked_tlbent();
 577
 578         tte_vaddr = (unsigned long) KERNBASE;
 579         phys_page = (prom_boot_mapping_phys_low >> 22UL) << 22UL;
 580         tte_data = kern_large_tte(phys_page);
 581
 582         kern_locked_tte_data = tte_data;
 583
 584         /* Now lock us into the TLBs via Hypervisor or OBP. */
 585         if (tlb_type == hypervisor) {
 586                 for (i = 0; i < num_kernel_image_mappings; i++) {
 587                         hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_DMMU);
 588                         hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_IMMU);
 589                         tte_vaddr += 0x400000;
 590                         tte_data += 0x400000;
 591                 }
 592         } else {
 593                 for (i = 0; i < num_kernel_image_mappings; i++) {
 594                         prom_dtlb_load(tlb_ent - i, tte_data, tte_vaddr);
 595                         prom_itlb_load(tlb_ent - i, tte_data, tte_vaddr);
 596                         tte_vaddr += 0x400000;
 597                         tte_data += 0x400000;
 598                 }
 599                 sparc64_highest_unlocked_tlb_ent = tlb_ent - i;
 600         }
 601         if (tlb_type == cheetah_plus) {
 602                 sparc64_kern_pri_context = (CTX_CHEETAH_PLUS_CTX0 |
 603                                             CTX_CHEETAH_PLUS_NUC);
 604                 sparc64_kern_pri_nuc_bits = CTX_CHEETAH_PLUS_NUC;
 605                 sparc64_kern_sec_context = CTX_CHEETAH_PLUS_CTX0;
 606         }
 607 }
 608
 609
 610 static void __init inherit_prom_mappings(void)
 611 {
 612         read_obp_translations();
 613
 614         /* Now fixup OBP's idea about where we really are mapped. */
 615         printk("Remapping the kernel... ");
 616         remap_kernel();
 617         printk("done.\n");
 618 }
 619
 620 void prom_world(int enter)
 621 {
 622         if (!enter)
 623                 set_fs((mm_segment_t) { get_thread_current_ds() });
 624
 625         __asm__ __volatile__("flushw");
 626 }
 627
 628 void __flush_dcache_range(unsigned long start, unsigned long end)
 629 {
 630         unsigned long va;
 631
 632         if (tlb_type == spitfire) {
 633                 int n = 0;
 634
 635                 for (va = start; va < end; va += 32) {
 636                         spitfire_put_dcache_tag(va & 0x3fe0, 0x0);
 637                         if (++n >= 512)
 638                                 break;
 639                 }
 640         } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
 641                 start = __pa(start);
 642                 end = __pa(end);
 643                 for (va = start; va < end; va += 32)
 644                         __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
 645                                              "membar #Sync"
 646                                              : /* no outputs */
 647                                              : "r" (va),
 648                                                "i" (ASI_DCACHE_INVALIDATE));
 649         }
 650 }
 651
 652 /* get_new_mmu_context() uses "cache + 1".  */
 653 DEFINE_SPINLOCK(ctx_alloc_lock);
 654 unsigned long tlb_context_cache = CTX_FIRST_VERSION - 1;
 655 #define MAX_CTX_NR      (1UL << CTX_NR_BITS)
 656 #define CTX_BMAP_SLOTS  BITS_TO_LONGS(MAX_CTX_NR)
 657 DECLARE_BITMAP(mmu_context_bmap, MAX_CTX_NR);
 658
 659 /* Caller does TLB context flushing on local CPU if necessary.
 660  * The caller also ensures that CTX_VALID(mm->context) is false.
 661  *
 662  * We must be careful about boundary cases so that we never
 663  * let the user have CTX 0 (nucleus) or we ever use a CTX
 664  * version of zero (and thus NO_CONTEXT would not be caught
 665  * by version mis-match tests in mmu_context.h).
 666  *
 667  * Always invoked with interrupts disabled.
 668  */
 669 void get_new_mmu_context(struct mm_struct *mm)
 670 {
 671         unsigned long ctx, new_ctx;
 672         unsigned long orig_pgsz_bits;
 673         unsigned long flags;
 674         int new_version;
 675
 676         spin_lock_irqsave(&ctx_alloc_lock, flags);
 677         orig_pgsz_bits = (mm->context.sparc64_ctx_val & CTX_PGSZ_MASK);
 678         ctx = (tlb_context_cache + 1) & CTX_NR_MASK;
 679         new_ctx = find_next_zero_bit(mmu_context_bmap, 1 << CTX_NR_BITS, ctx);
 680         new_version = 0;
 681         if (new_ctx >= (1 << CTX_NR_BITS)) {
 682                 new_ctx = find_next_zero_bit(mmu_context_bmap, ctx, 1);
 683                 if (new_ctx >= ctx) {
 684                         int i;
 685                         new_ctx = (tlb_context_cache & CTX_VERSION_MASK) +
 686                                 CTX_FIRST_VERSION;
 687                         if (new_ctx == 1)
 688                                 new_ctx = CTX_FIRST_VERSION;
 689
 690                         /* Don't call memset, for 16 entries that's just
 691                          * plain silly...
 692                          */
 693                         mmu_context_bmap[0] = 3;
 694                         mmu_context_bmap[1] = 0;
 695                         mmu_context_bmap[2] = 0;
 696                         mmu_context_bmap[3] = 0;
 697                         for (i = 4; i < CTX_BMAP_SLOTS; i += 4) {
 698                                 mmu_context_bmap[i + 0] = 0;
 699                                 mmu_context_bmap[i + 1] = 0;
 700                                 mmu_context_bmap[i + 2] = 0;
 701                                 mmu_context_bmap[i + 3] = 0;
 702                         }
 703                         new_version = 1;
 704                         goto out;
 705                 }
 706         }
 707         mmu_context_bmap[new_ctx>>6] |= (1UL << (new_ctx & 63));
 708         new_ctx |= (tlb_context_cache & CTX_VERSION_MASK);
 709 out:
 710         tlb_context_cache = new_ctx;
 711         mm->context.sparc64_ctx_val = new_ctx | orig_pgsz_bits;
 712         spin_unlock_irqrestore(&ctx_alloc_lock, flags);
 713
 714         if (unlikely(new_version))
 715                 smp_new_mmu_context_version();
 716 }
 717
 718 static int numa_enabled = 1;
 719 static int numa_debug;
 720
 721 static int __init early_numa(char *p)
 722 {
 723         if (!p)
 724                 return 0;
 725
 726         if (strstr(p, "off"))
 727                 numa_enabled = 0;
 728
 729         if (strstr(p, "debug"))
 730                 numa_debug = 1;
 731
 732         return 0;
 733 }
 734 early_param("numa", early_numa);
 735
 736 #define numadbg(f, a...) \
 737 do {    if (numa_debug) \
 738                 printk(KERN_INFO f, ## a); \
 739 } while (0)
 740
 741 static void __init find_ramdisk(unsigned long phys_base)
 742 {
 743 #ifdef CONFIG_BLK_DEV_INITRD
 744         if (sparc_ramdisk_image || sparc_ramdisk_image64) {
 745                 unsigned long ramdisk_image;
 746
 747                 /* Older versions of the bootloader only supported a
 748                  * 32-bit physical address for the ramdisk image
 749                  * location, stored at sparc_ramdisk_image.  Newer
 750                  * SILO versions set sparc_ramdisk_image to zero and
 751                  * provide a full 64-bit physical address at
 752                  * sparc_ramdisk_image64.
 753                  */
 754                 ramdisk_image = sparc_ramdisk_image;
 755                 if (!ramdisk_image)
 756                         ramdisk_image = sparc_ramdisk_image64;
 757
 758                 /* Another bootloader quirk.  The bootloader normalizes
 759                  * the physical address to KERNBASE, so we have to
 760                  * factor that back out and add in the lowest valid
 761                  * physical page address to get the true physical address.
 762                  */
 763                 ramdisk_image -= KERNBASE;
 764                 ramdisk_image += phys_base;
 765
 766                 numadbg("Found ramdisk at physical address 0x%lx, size %u\n",
 767                         ramdisk_image, sparc_ramdisk_size);
 768
 769                 initrd_start = ramdisk_image;
 770                 initrd_end = ramdisk_image + sparc_ramdisk_size;
 771
 772                 lmb_reserve(initrd_start, initrd_end);
 773         }
 774 #endif
 775 }
 776
 777 struct node_mem_mask {
 778         unsigned long mask;
 779         unsigned long val;
 780         unsigned long bootmem_paddr;
 781 };
 782 static struct node_mem_mask node_masks[MAX_NUMNODES];
 783 static int num_node_masks;
 784
 785 int numa_cpu_lookup_table[NR_CPUS];
 786 cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
 787
 788 #ifdef CONFIG_NEED_MULTIPLE_NODES
 789 static bootmem_data_t plat_node_bdata[MAX_NUMNODES];
 790
 791 struct mdesc_mblock {
 792         u64     base;
 793         u64     size;
 794         u64     offset; /* RA-to-PA */
 795 };
 796 static struct mdesc_mblock *mblocks;
 797 static int num_mblocks;
 798
 799 static unsigned long ra_to_pa(unsigned long addr)
 800 {
 801         int i;
 802
 803         for (i = 0; i < num_mblocks; i++) {
 804                 struct mdesc_mblock *m = &mblocks[i];
 805
 806                 if (addr >= m->base &&
 807                     addr < (m->base + m->size)) {
 808                         addr += m->offset;
 809                         break;
 810                 }
 811         }
 812         return addr;
 813 }
 814
 815 static int find_node(unsigned long addr)
 816 {
 817         int i;
 818
 819         addr = ra_to_pa(addr);
 820         for (i = 0; i < num_node_masks; i++) {
 821                 struct node_mem_mask *p = &node_masks[i];
 822
 823                 if ((addr & p->mask) == p->val)
 824                         return i;
 825         }
 826         return -1;
 827 }
 828
 829 static unsigned long nid_range(unsigned long start, unsigned long end,
 830                                int *nid)
 831 {
 832         *nid = find_node(start);
 833         start += PAGE_SIZE;
 834         while (start < end) {
 835                 int n = find_node(start);
 836
 837                 if (n != *nid)
 838                         break;
 839                 start += PAGE_SIZE;
 840         }
 841
 842         return start;
 843 }
 844 #else
 845 static unsigned long nid_range(unsigned long start, unsigned long end,
 846                                int *nid)
 847 {
 848         *nid = 0;
 849         return end;
 850 }
 851 #endif
 852
 853 /* This must be invoked after performing all of the necessary
 854  * add_active_range() calls for 'nid'.  We need to be able to get
 855  * correct data from get_pfn_range_for_nid().
 856  */
 857 static void __init allocate_node_data(int nid)
 858 {
 859         unsigned long paddr, num_pages, start_pfn, end_pfn;
 860         struct pglist_data *p;
 861
 862 #ifdef CONFIG_NEED_MULTIPLE_NODES
 863         paddr = lmb_alloc_nid(sizeof(struct pglist_data),
 864                               SMP_CACHE_BYTES, nid, nid_range);
 865         if (!paddr) {
 866                 prom_printf("Cannot allocate pglist_data for nid[%d]\n", nid);
 867                 prom_halt();
 868         }
 869         NODE_DATA(nid) = __va(paddr);
 870         memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
 871
 872         NODE_DATA(nid)->bdata = &plat_node_bdata[nid];
 873 #endif
 874
 875         p = NODE_DATA(nid);
 876
 877         get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
 878         p->node_start_pfn = start_pfn;
 879         p->node_spanned_pages = end_pfn - start_pfn;
 880
 881         if (p->node_spanned_pages) {
 882                 num_pages = bootmem_bootmap_pages(p->node_spanned_pages);
 883
 884                 paddr = lmb_alloc_nid(num_pages << PAGE_SHIFT, PAGE_SIZE, nid,
 885                                       nid_range);
 886                 if (!paddr) {
 887                         prom_printf("Cannot allocate bootmap for nid[%d]\n",
 888                                   nid);
 889                         prom_halt();
 890                 }
 891                 node_masks[nid].bootmem_paddr = paddr;
 892         }
 893 }
 894
 895 static void init_node_masks_nonnuma(void)
 896 {
 897         int i;
 898
 899         numadbg("Initializing tables for non-numa.\n");
 900
 901         node_masks[0].mask = node_masks[0].val = 0;
 902         num_node_masks = 1;
 903
 904         for (i = 0; i < NR_CPUS; i++)
 905                 numa_cpu_lookup_table[i] = 0;
 906
 907         numa_cpumask_lookup_table[0] = CPU_MASK_ALL;
 908 }
 909
 910 #ifdef CONFIG_NEED_MULTIPLE_NODES
 911 struct pglist_data *node_data[MAX_NUMNODES];
 912
 913 EXPORT_SYMBOL(numa_cpu_lookup_table);
 914 EXPORT_SYMBOL(numa_cpumask_lookup_table);
 915 EXPORT_SYMBOL(node_data);
 916
 917 struct mdesc_mlgroup {
 918         u64     node;
 919         u64     latency;
 920         u64     match;
 921         u64     mask;
 922 };
 923 static struct mdesc_mlgroup *mlgroups;
 924 static int num_mlgroups;
 925
 926 static int scan_pio_for_cfg_handle(struct mdesc_handle *md, u64 pio,
 927                                    u32 cfg_handle)
 928 {
 929         u64 arc;
 930
 931         mdesc_for_each_arc(arc, md, pio, MDESC_ARC_TYPE_FWD) {
 932                 u64 target = mdesc_arc_target(md, arc);
 933                 const u64 *val;
 934
 935                 val = mdesc_get_property(md, target,
 936                                          "cfg-handle", NULL);
 937                 if (val && *val == cfg_handle)
 938                         return 0;
 939         }
 940         return -ENODEV;
 941 }
 942
 943 static int scan_arcs_for_cfg_handle(struct mdesc_handle *md, u64 grp,
 944                                     u32 cfg_handle)
 945 {
 946         u64 arc, candidate, best_latency = ~(u64)0;
 947
 948         candidate = MDESC_NODE_NULL;
 949         mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
 950                 u64 target = mdesc_arc_target(md, arc);
 951                 const char *name = mdesc_node_name(md, target);
 952                 const u64 *val;
 953
 954                 if (strcmp(name, "pio-latency-group"))
 955                         continue;
 956
 957                 val = mdesc_get_property(md, target, "latency", NULL);
 958                 if (!val)
 959                         continue;
 960
 961                 if (*val < best_latency) {
 962                         candidate = target;
 963                         best_latency = *val;
 964                 }
 965         }
 966
 967         if (candidate == MDESC_NODE_NULL)
 968                 return -ENODEV;
 969
 970         return scan_pio_for_cfg_handle(md, candidate, cfg_handle);
 971 }
 972
 973 int of_node_to_nid(struct device_node *dp)
 974 {
 975         const struct linux_prom64_registers *regs;
 976         struct mdesc_handle *md;
 977         u32 cfg_handle;
 978         int count, nid;
 979         u64 grp;
 980
 981         if (!mlgroups)
 982                 return -1;
 983
 984         regs = of_get_property(dp, "reg", NULL);
 985         if (!regs)
 986                 return -1;
 987
 988         cfg_handle = (regs->phys_addr >> 32UL) & 0x0fffffff;
 989
 990         md = mdesc_grab();
 991
 992         count = 0;
 993         nid = -1;
 994         mdesc_for_each_node_by_name(md, grp, "group") {
 995                 if (!scan_arcs_for_cfg_handle(md, grp, cfg_handle)) {
 996                         nid = count;
 997                         break;
 998                 }
 999                 count++;
1000         }
1001
1002         mdesc_release(md);
1003
1004         return nid;
1005 }
1006
1007 static void add_node_ranges(void)
1008 {
1009         int i;
1010
1011         for (i = 0; i < lmb.memory.cnt; i++) {
1012                 unsigned long size = lmb_size_bytes(&lmb.memory, i);
1013                 unsigned long start, end;
1014
1015                 start = lmb.memory.region[i].base;
1016                 end = start + size;
1017                 while (start < end) {
1018                         unsigned long this_end;
1019                         int nid;
1020
1021                         this_end = nid_range(start, end, &nid);
1022
1023                         numadbg("Adding active range nid[%d] "
1024                                 "start[%lx] end[%lx]\n",
1025                                 nid, start, this_end);
1026
1027                         add_active_range(nid,
1028                                          start >> PAGE_SHIFT,
1029                                          this_end >> PAGE_SHIFT);
1030
1031                         start = this_end;
1032                 }
1033         }
1034 }
1035
1036 static int __init grab_mlgroups(struct mdesc_handle *md)
1037 {
1038         unsigned long paddr;
1039         int count = 0;
1040         u64 node;
1041
1042         mdesc_for_each_node_by_name(md, node, "memory-latency-group")
1043                 count++;
1044         if (!count)
1045                 return -ENOENT;
1046
1047         paddr = lmb_alloc(count * sizeof(struct mdesc_mlgroup),
1048                           SMP_CACHE_BYTES);
1049         if (!paddr)
1050                 return -ENOMEM;
1051
1052         mlgroups = __va(paddr);
1053         num_mlgroups = count;
1054
1055         count = 0;
1056         mdesc_for_each_node_by_name(md, node, "memory-latency-group") {
1057                 struct mdesc_mlgroup *m = &mlgroups[count++];
1058                 const u64 *val;
1059
1060                 m->node = node;
1061
1062                 val = mdesc_get_property(md, node, "latency", NULL);
1063                 m->latency = *val;
1064                 val = mdesc_get_property(md, node, "address-match", NULL);
1065                 m->match = *val;
1066                 val = mdesc_get_property(md, node, "address-mask", NULL);
1067                 m->mask = *val;
1068
1069                 numadbg("MLGROUP[%d]: node[%lx] latency[%lx] "
1070                         "match[%lx] mask[%lx]\n",
1071                         count - 1, m->node, m->latency, m->match, m->mask);
1072         }
1073
1074         return 0;
1075 }
1076
1077 static int __init grab_mblocks(struct mdesc_handle *md)
1078 {
1079         unsigned long paddr;
1080         int count = 0;
1081         u64 node;
1082
1083         mdesc_for_each_node_by_name(md, node, "mblock")
1084                 count++;
1085         if (!count)
1086                 return -ENOENT;
1087
1088         paddr = lmb_alloc(count * sizeof(struct mdesc_mblock),
1089                           SMP_CACHE_BYTES);
1090         if (!paddr)
1091                 return -ENOMEM;
1092
1093         mblocks = __va(paddr);
1094         num_mblocks = count;
1095
1096         count = 0;
1097         mdesc_for_each_node_by_name(md, node, "mblock") {
1098                 struct mdesc_mblock *m = &mblocks[count++];
1099                 const u64 *val;
1100
1101                 val = mdesc_get_property(md, node, "base", NULL);
1102                 m->base = *val;
1103                 val = mdesc_get_property(md, node, "size", NULL);
1104                 m->size = *val;
1105                 val = mdesc_get_property(md, node,
1106                                          "address-congruence-offset", NULL);
1107                 m->offset = *val;
1108
1109                 numadbg("MBLOCK[%d]: base[%lx] size[%lx] offset[%lx]\n",
1110                         count - 1, m->base, m->size, m->offset);
1111         }
1112
1113         return 0;
1114 }
1115
1116 static void __init numa_parse_mdesc_group_cpus(struct mdesc_handle *md,
1117                                                u64 grp, cpumask_t *mask)
1118 {
1119         u64 arc;
1120
1121         cpus_clear(*mask);
1122
1123         mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_BACK) {
1124                 u64 target = mdesc_arc_target(md, arc);
1125                 const char *name = mdesc_node_name(md, target);
1126                 const u64 *id;
1127
1128                 if (strcmp(name, "cpu"))
1129                         continue;
1130                 id = mdesc_get_property(md, target, "id", NULL);
1131                 if (*id < NR_CPUS)
1132                         cpu_set(*id, *mask);
1133         }
1134 }
1135
1136 static struct mdesc_mlgroup * __init find_mlgroup(u64 node)
1137 {
1138         int i;
1139
1140         for (i = 0; i < num_mlgroups; i++) {
1141                 struct mdesc_mlgroup *m = &mlgroups[i];
1142                 if (m->node == node)
1143                         return m;
1144         }
1145         return NULL;
1146 }
1147
1148 static int __init numa_attach_mlgroup(struct mdesc_handle *md, u64 grp,
1149                                       int index)
1150 {
1151         struct mdesc_mlgroup *candidate = NULL;
1152         u64 arc, best_latency = ~(u64)0;
1153         struct node_mem_mask *n;
1154
1155         mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1156                 u64 target = mdesc_arc_target(md, arc);
1157                 struct mdesc_mlgroup *m = find_mlgroup(target);
1158                 if (!m)
1159                         continue;
1160                 if (m->latency < best_latency) {
1161                         candidate = m;
1162                         best_latency = m->latency;
1163                 }
1164         }
1165         if (!candidate)
1166                 return -ENOENT;
1167
1168         if (num_node_masks != index) {
1169                 printk(KERN_ERR "Inconsistent NUMA state, "
1170                        "index[%d] != num_node_masks[%d]\n",
1171                        index, num_node_masks);
1172                 return -EINVAL;
1173         }
1174
1175         n = &node_masks[num_node_masks++];
1176
1177         n->mask = candidate->mask;
1178         n->val = candidate->match;
1179
1180         numadbg("NUMA NODE[%d]: mask[%lx] val[%lx] (latency[%lx])\n",
1181                 index, n->mask, n->val, candidate->latency);
1182
1183         return 0;
1184 }
1185
1186 static int __init numa_parse_mdesc_group(struct mdesc_handle *md, u64 grp,
1187                                          int index)
1188 {
1189         cpumask_t mask;
1190         int cpu;
1191
1192         numa_parse_mdesc_group_cpus(md, grp, &mask);
1193
1194         for_each_cpu_mask(cpu, mask)
1195                 numa_cpu_lookup_table[cpu] = index;
1196         numa_cpumask_lookup_table[index] = mask;
1197
1198         if (numa_debug) {
1199                 printk(KERN_INFO "NUMA GROUP[%d]: cpus [ ", index);
1200                 for_each_cpu_mask(cpu, mask)
1201                         printk("%d ", cpu);
1202                 printk("]\n");
1203         }
1204
1205         return numa_attach_mlgroup(md, grp, index);
1206 }
1207
1208 static int __init numa_parse_mdesc(void)
1209 {
1210         struct mdesc_handle *md = mdesc_grab();
1211         int i, err, count;
1212         u64 node;
1213
1214         node = mdesc_node_by_name(md, MDESC_NODE_NULL, "latency-groups");
1215         if (node == MDESC_NODE_NULL) {
1216                 mdesc_release(md);
1217                 return -ENOENT;
1218         }
1219
1220         err = grab_mblocks(md);
1221         if (err < 0)
1222                 goto out;
1223
1224         err = grab_mlgroups(md);
1225         if (err < 0)
1226                 goto out;
1227
1228         count = 0;
1229         mdesc_for_each_node_by_name(md, node, "group") {
1230                 err = numa_parse_mdesc_group(md, node, count);
1231                 if (err < 0)
1232                         break;
1233                 count++;
1234         }
1235
1236         add_node_ranges();
1237
1238         for (i = 0; i < num_node_masks; i++) {
1239                 allocate_node_data(i);
1240                 node_set_online(i);
1241         }
1242
1243         err = 0;
1244 out:
1245         mdesc_release(md);
1246         return err;
1247 }
1248
1249 static int __init numa_parse_sun4u(void)
1250 {
1251         return -1;
1252 }
1253
1254 static int __init bootmem_init_numa(void)
1255 {
1256         int err = -1;
1257
1258         numadbg("bootmem_init_numa()\n");
1259
1260         if (numa_enabled) {
1261                 if (tlb_type == hypervisor)
1262                         err = numa_parse_mdesc();
1263                 else
1264                         err = numa_parse_sun4u();
1265         }
1266         return err;
1267 }
1268
1269 #else
1270
1271 static int bootmem_init_numa(void)
1272 {
1273         return -1;
1274 }
1275
1276 #endif
1277
1278 static void __init bootmem_init_nonnuma(void)
1279 {
1280         unsigned long top_of_ram = lmb_end_of_DRAM();
1281         unsigned long total_ram = lmb_phys_mem_size();
1282         unsigned int i;
1283
1284         numadbg("bootmem_init_nonnuma()\n");
1285
1286         printk(KERN_INFO "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
1287                top_of_ram, total_ram);
1288         printk(KERN_INFO "Memory hole size: %ldMB\n",
1289                (top_of_ram - total_ram) >> 20);
1290
1291         init_node_masks_nonnuma();
1292
1293         for (i = 0; i < lmb.memory.cnt; i++) {
1294                 unsigned long size = lmb_size_bytes(&lmb.memory, i);
1295                 unsigned long start_pfn, end_pfn;
1296
1297                 if (!size)
1298                         continue;
1299
1300                 start_pfn = lmb.memory.region[i].base >> PAGE_SHIFT;
1301                 end_pfn = start_pfn + lmb_size_pages(&lmb.memory, i);
1302                 add_active_range(0, start_pfn, end_pfn);
1303         }
1304
1305         allocate_node_data(0);
1306
1307         node_set_online(0);
1308 }
1309
1310 static void __init reserve_range_in_node(int nid, unsigned long start,
1311                                          unsigned long end)
1312 {
1313         numadbg("    reserve_range_in_node(nid[%d],start[%lx],end[%lx]\n",
1314                 nid, start, end);
1315         while (start < end) {
1316                 unsigned long this_end;
1317                 int n;
1318
1319                 this_end = nid_range(start, end, &n);
1320                 if (n == nid) {
1321                         numadbg("      MATCH reserving range [%lx:%lx]\n",
1322                                 start, this_end);
1323                         reserve_bootmem_node(NODE_DATA(nid), start,
1324                                              (this_end - start), BOOTMEM_DEFAULT);
1325                 } else
1326                         numadbg("      NO MATCH, advancing start to %lx\n",
1327                                 this_end);
1328
1329                 start = this_end;
1330         }
1331 }
1332
1333 static void __init trim_reserved_in_node(int nid)
1334 {
1335         int i;
1336
1337         numadbg("  trim_reserved_in_node(%d)\n", nid);
1338
1339         for (i = 0; i < lmb.reserved.cnt; i++) {
1340                 unsigned long start = lmb.reserved.region[i].base;
1341                 unsigned long size = lmb_size_bytes(&lmb.reserved, i);
1342                 unsigned long end = start + size;
1343
1344                 reserve_range_in_node(nid, start, end);
1345         }
1346 }
1347
1348 static void __init bootmem_init_one_node(int nid)
1349 {
1350         struct pglist_data *p;
1351
1352         numadbg("bootmem_init_one_node(%d)\n", nid);
1353
1354         p = NODE_DATA(nid);
1355
1356         if (p->node_spanned_pages) {
1357                 unsigned long paddr = node_masks[nid].bootmem_paddr;
1358                 unsigned long end_pfn;
1359
1360                 end_pfn = p->node_start_pfn + p->node_spanned_pages;
1361
1362                 numadbg("  init_bootmem_node(%d, %lx, %lx, %lx)\n",
1363                         nid, paddr >> PAGE_SHIFT, p->node_start_pfn, end_pfn);
1364
1365                 init_bootmem_node(p, paddr >> PAGE_SHIFT,
1366                                   p->node_start_pfn, end_pfn);
1367
1368                 numadbg("  free_bootmem_with_active_regions(%d, %lx)\n",
1369                         nid, end_pfn);
1370                 free_bootmem_with_active_regions(nid, end_pfn);
1371
1372                 trim_reserved_in_node(nid);
1373
1374                 numadbg("  sparse_memory_present_with_active_regions(%d)\n",
1375                         nid);
1376                 sparse_memory_present_with_active_regions(nid);
1377         }
1378 }
1379
1380 static unsigned long __init bootmem_init(unsigned long phys_base)
1381 {
1382         unsigned long end_pfn;
1383         int nid;
1384
1385         end_pfn = lmb_end_of_DRAM() >> PAGE_SHIFT;
1386         max_pfn = max_low_pfn = end_pfn;
1387         min_low_pfn = (phys_base >> PAGE_SHIFT);
1388
1389         if (bootmem_init_numa() < 0)
1390                 bootmem_init_nonnuma();
1391
1392         /* XXX cpu notifier XXX */
1393
1394         for_each_online_node(nid)
1395                 bootmem_init_one_node(nid);
1396
1397         sparse_init();
1398
1399         return end_pfn;
1400 }
1401
1402 static struct linux_prom64_registers pall[MAX_BANKS] __initdata;
1403 static int pall_ents __initdata;
1404
1405 #ifdef CONFIG_DEBUG_PAGEALLOC
1406 static unsigned long __ref kernel_map_range(unsigned long pstart,
1407                                             unsigned long pend, pgprot_t prot)
1408 {
1409         unsigned long vstart = PAGE_OFFSET + pstart;
1410         unsigned long vend = PAGE_OFFSET + pend;
1411         unsigned long alloc_bytes = 0UL;
1412
1413         if ((vstart & ~PAGE_MASK) || (vend & ~PAGE_MASK)) {
1414                 prom_printf("kernel_map: Unaligned physmem[%lx:%lx]\n",
1415                             vstart, vend);
1416                 prom_halt();
1417         }
1418
1419         while (vstart < vend) {
1420                 unsigned long this_end, paddr = __pa(vstart);
1421                 pgd_t *pgd = pgd_offset_k(vstart);
1422                 pud_t *pud;
1423                 pmd_t *pmd;
1424                 pte_t *pte;
1425
1426                 pud = pud_offset(pgd, vstart);
1427                 if (pud_none(*pud)) {
1428                         pmd_t *new;
1429
1430                         new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1431                         alloc_bytes += PAGE_SIZE;
1432                         pud_populate(&init_mm, pud, new);
1433                 }
1434
1435                 pmd = pmd_offset(pud, vstart);
1436                 if (!pmd_present(*pmd)) {
1437                         pte_t *new;
1438
1439                         new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1440                         alloc_bytes += PAGE_SIZE;
1441                         pmd_populate_kernel(&init_mm, pmd, new);
1442                 }
1443
1444                 pte = pte_offset_kernel(pmd, vstart);
1445                 this_end = (vstart + PMD_SIZE) & PMD_MASK;
1446                 if (this_end > vend)
1447                         this_end = vend;
1448
1449                 while (vstart < this_end) {
1450                         pte_val(*pte) = (paddr | pgprot_val(prot));
1451
1452                         vstart += PAGE_SIZE;
1453                         paddr += PAGE_SIZE;
1454                         pte++;
1455                 }
1456         }
1457
1458         return alloc_bytes;
1459 }
1460
1461 extern unsigned int kvmap_linear_patch[1];
1462 #endif /* CONFIG_DEBUG_PAGEALLOC */
1463
1464 static void __init mark_kpte_bitmap(unsigned long start, unsigned long end)
1465 {
1466         const unsigned long shift_256MB = 28;
1467         const unsigned long mask_256MB = ((1UL << shift_256MB) - 1UL);
1468         const unsigned long size_256MB = (1UL << shift_256MB);
1469
1470         while (start < end) {
1471                 long remains;
1472
1473                 remains = end - start;
1474                 if (remains < size_256MB)
1475                         break;
1476
1477                 if (start & mask_256MB) {
1478                         start = (start + size_256MB) & ~mask_256MB;
1479                         continue;
1480                 }
1481
1482                 while (remains >= size_256MB) {
1483                         unsigned long index = start >> shift_256MB;
1484
1485                         __set_bit(index, kpte_linear_bitmap);
1486
1487                         start += size_256MB;
1488                         remains -= size_256MB;
1489                 }
1490         }
1491 }
1492
1493 static void __init init_kpte_bitmap(void)
1494 {
1495         unsigned long i;
1496
1497         for (i = 0; i < pall_ents; i++) {
1498                 unsigned long phys_start, phys_end;
1499
1500                 phys_start = pall[i].phys_addr;
1501                 phys_end = phys_start + pall[i].reg_size;
1502
1503                 mark_kpte_bitmap(phys_start, phys_end);
1504         }
1505 }
1506
1507 static void __init kernel_physical_mapping_init(void)
1508 {
1509 #ifdef CONFIG_DEBUG_PAGEALLOC
1510         unsigned long i, mem_alloced = 0UL;
1511
1512         for (i = 0; i < pall_ents; i++) {
1513                 unsigned long phys_start, phys_end;
1514
1515                 phys_start = pall[i].phys_addr;
1516                 phys_end = phys_start + pall[i].reg_size;
1517
1518                 mem_alloced += kernel_map_range(phys_start, phys_end,
1519                                                 PAGE_KERNEL);
1520         }
1521
1522         printk("Allocated %ld bytes for kernel page tables.\n",
1523                mem_alloced);
1524
1525         kvmap_linear_patch[0] = 0x01000000; /* nop */
1526         flushi(&kvmap_linear_patch[0]);
1527
1528         __flush_tlb_all();
1529 #endif
1530 }
1531
1532 #ifdef CONFIG_DEBUG_PAGEALLOC
1533 void kernel_map_pages(struct page *page, int numpages, int enable)
1534 {
1535         unsigned long phys_start = page_to_pfn(page) << PAGE_SHIFT;
1536         unsigned long phys_end = phys_start + (numpages * PAGE_SIZE);
1537
1538         kernel_map_range(phys_start, phys_end,
1539                          (enable ? PAGE_KERNEL : __pgprot(0)));
1540
1541         flush_tsb_kernel_range(PAGE_OFFSET + phys_start,
1542                                PAGE_OFFSET + phys_end);
1543
1544         /* we should perform an IPI and flush all tlbs,
1545          * but that can deadlock->flush only current cpu.
1546          */
1547         __flush_tlb_kernel_range(PAGE_OFFSET + phys_start,
1548                                  PAGE_OFFSET + phys_end);
1549 }
1550 #endif
1551
1552 unsigned long __init find_ecache_flush_span(unsigned long size)
1553 {
1554         int i;
1555
1556         for (i = 0; i < pavail_ents; i++) {
1557                 if (pavail[i].reg_size >= size)
1558                         return pavail[i].phys_addr;
1559         }
1560
1561         return ~0UL;
1562 }
1563
1564 static void __init tsb_phys_patch(void)
1565 {
1566         struct tsb_ldquad_phys_patch_entry *pquad;
1567         struct tsb_phys_patch_entry *p;
1568
1569         pquad = &__tsb_ldquad_phys_patch;
1570         while (pquad < &__tsb_ldquad_phys_patch_end) {
1571                 unsigned long addr = pquad->addr;
1572
1573                 if (tlb_type == hypervisor)
1574                         *(unsigned int *) addr = pquad->sun4v_insn;
1575                 else
1576                         *(unsigned int *) addr = pquad->sun4u_insn;
1577                 wmb();
1578                 __asm__ __volatile__("flush     %0"
1579                                      : /* no outputs */
1580                                      : "r" (addr));
1581
1582                 pquad++;
1583         }
1584
1585         p = &__tsb_phys_patch;
1586         while (p < &__tsb_phys_patch_end) {
1587                 unsigned long addr = p->addr;
1588
1589                 *(unsigned int *) addr = p->insn;
1590                 wmb();
1591                 __asm__ __volatile__("flush     %0"
1592                                      : /* no outputs */
1593                                      : "r" (addr));
1594
1595                 p++;
1596         }
1597 }
1598
1599 /* Don't mark as init, we give this to the Hypervisor.  */
1600 #ifndef CONFIG_DEBUG_PAGEALLOC
1601 #define NUM_KTSB_DESCR  2
1602 #else
1603 #define NUM_KTSB_DESCR  1
1604 #endif
1605 static struct hv_tsb_descr ktsb_descr[NUM_KTSB_DESCR];
1606 extern struct tsb swapper_tsb[KERNEL_TSB_NENTRIES];
1607
1608 static void __init sun4v_ktsb_init(void)
1609 {
1610         unsigned long ktsb_pa;
1611
1612         /* First KTSB for PAGE_SIZE mappings.  */
1613         ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
1614
1615         switch (PAGE_SIZE) {
1616         case 8 * 1024:
1617         default:
1618                 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_8K;
1619                 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_8K;
1620                 break;
1621
1622         case 64 * 1024:
1623                 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_64K;
1624                 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_64K;
1625                 break;
1626
1627         case 512 * 1024:
1628                 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_512K;
1629                 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_512K;
1630                 break;
1631
1632         case 4 * 1024 * 1024:
1633                 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_4MB;
1634                 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_4MB;
1635                 break;
1636         };
1637
1638         ktsb_descr[0].assoc = 1;
1639         ktsb_descr[0].num_ttes = KERNEL_TSB_NENTRIES;
1640         ktsb_descr[0].ctx_idx = 0;
1641         ktsb_descr[0].tsb_base = ktsb_pa;
1642         ktsb_descr[0].resv = 0;
1643
1644 #ifndef CONFIG_DEBUG_PAGEALLOC
1645         /* Second KTSB for 4MB/256MB mappings.  */
1646         ktsb_pa = (kern_base +
1647                    ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
1648
1649         ktsb_descr[1].pgsz_idx = HV_PGSZ_IDX_4MB;
1650         ktsb_descr[1].pgsz_mask = (HV_PGSZ_MASK_4MB |
1651                                    HV_PGSZ_MASK_256MB);
1652         ktsb_descr[1].assoc = 1;
1653         ktsb_descr[1].num_ttes = KERNEL_TSB4M_NENTRIES;
1654         ktsb_descr[1].ctx_idx = 0;
1655         ktsb_descr[1].tsb_base = ktsb_pa;
1656         ktsb_descr[1].resv = 0;
1657 #endif
1658 }
1659
1660 void __cpuinit sun4v_ktsb_register(void)
1661 {
1662         unsigned long pa, ret;
1663
1664         pa = kern_base + ((unsigned long)&ktsb_descr[0] - KERNBASE);
1665
1666         ret = sun4v_mmu_tsb_ctx0(NUM_KTSB_DESCR, pa);
1667         if (ret != 0) {
1668                 prom_printf("hypervisor_mmu_tsb_ctx0[%lx]: "
1669                             "errors with %lx\n", pa, ret);
1670                 prom_halt();
1671         }
1672 }
1673
1674 /* paging_init() sets up the page tables */
1675
1676 extern void central_probe(void);
1677
1678 static unsigned long last_valid_pfn;
1679 pgd_t swapper_pg_dir[2048];
1680
1681 static void sun4u_pgprot_init(void);
1682 static void sun4v_pgprot_init(void);
1683
1684 /* Dummy function */
1685 void __init setup_per_cpu_areas(void)
1686 {
1687 }
1688
1689 void __init paging_init(void)
1690 {
1691         unsigned long end_pfn, shift, phys_base;
1692         unsigned long real_end, i;
1693
1694         /* These build time checkes make sure that the dcache_dirty_cpu()
1695          * page->flags usage will work.
1696          *
1697          * When a page gets marked as dcache-dirty, we store the
1698          * cpu number starting at bit 32 in the page->flags.  Also,
1699          * functions like clear_dcache_dirty_cpu use the cpu mask
1700          * in 13-bit signed-immediate instruction fields.
1701          */
1702         BUILD_BUG_ON(FLAGS_RESERVED != 32);
1703         BUILD_BUG_ON(SECTIONS_WIDTH + NODES_WIDTH + ZONES_WIDTH +
1704                      ilog2(roundup_pow_of_two(NR_CPUS)) > FLAGS_RESERVED);
1705         BUILD_BUG_ON(NR_CPUS > 4096);
1706
1707         kern_base = (prom_boot_mapping_phys_low >> 22UL) << 22UL;
1708         kern_size = (unsigned long)&_end - (unsigned long)KERNBASE;
1709
1710         sstate_booting();
1711
1712         /* Invalidate both kernel TSBs.  */
1713         memset(swapper_tsb, 0x40, sizeof(swapper_tsb));
1714 #ifndef CONFIG_DEBUG_PAGEALLOC
1715         memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
1716 #endif
1717
1718         if (tlb_type == hypervisor)
1719                 sun4v_pgprot_init();
1720         else
1721                 sun4u_pgprot_init();
1722
1723         if (tlb_type == cheetah_plus ||
1724             tlb_type == hypervisor)
1725                 tsb_phys_patch();
1726
1727         if (tlb_type == hypervisor) {
1728                 sun4v_patch_tlb_handlers();
1729                 sun4v_ktsb_init();
1730         }
1731
1732         lmb_init();
1733
1734         /* Find available physical memory... */
1735         read_obp_memory("available", &pavail[0], &pavail_ents);
1736
1737         phys_base = 0xffffffffffffffffUL;
1738         for (i = 0; i < pavail_ents; i++) {
1739                 phys_base = min(phys_base, pavail[i].phys_addr);
1740                 lmb_add(pavail[i].phys_addr, pavail[i].reg_size);
1741         }
1742
1743         lmb_reserve(kern_base, kern_size);
1744
1745         find_ramdisk(phys_base);
1746
1747         if (cmdline_memory_size)
1748                 lmb_enforce_memory_limit(phys_base + cmdline_memory_size);
1749
1750         lmb_analyze();
1751         lmb_dump_all();
1752
1753         set_bit(0, mmu_context_bmap);
1754
1755         shift = kern_base + PAGE_OFFSET - ((unsigned long)KERNBASE);
1756
1757         real_end = (unsigned long)_end;
1758         num_kernel_image_mappings = DIV_ROUND_UP(real_end - KERNBASE, 1 << 22);
1759         printk("Kernel: Using %d locked TLB entries for main kernel image.\n",
1760                num_kernel_image_mappings);
1761
1762         /* Set kernel pgd to upper alias so physical page computations
1763          * work.
1764          */
1765         init_mm.pgd += ((shift) / (sizeof(pgd_t)));
1766
1767         memset(swapper_low_pmd_dir, 0, sizeof(swapper_low_pmd_dir));
1768
1769         /* Now can init the kernel/bad page tables. */
1770         pud_set(pud_offset(&swapper_pg_dir[0], 0),
1771                 swapper_low_pmd_dir + (shift / sizeof(pgd_t)));
1772
1773         inherit_prom_mappings();
1774
1775         read_obp_memory("reg", &pall[0], &pall_ents);
1776
1777         init_kpte_bitmap();
1778
1779         /* Ok, we can use our TLB miss and window trap handlers safely.  */
1780         setup_tba();
1781
1782         __flush_tlb_all();
1783
1784         if (tlb_type == hypervisor)
1785                 sun4v_ktsb_register();
1786
1787         /* We must setup the per-cpu areas before we pull in the
1788          * PROM and the MDESC.  The code there fills in cpu and
1789          * other information into per-cpu data structures.
1790          */
1791         real_setup_per_cpu_areas();
1792
1793         prom_build_devicetree();
1794
1795         if (tlb_type == hypervisor)
1796                 sun4v_mdesc_init();
1797
1798         /* Setup bootmem... */
1799         last_valid_pfn = end_pfn = bootmem_init(phys_base);
1800
1801 #ifndef CONFIG_NEED_MULTIPLE_NODES
1802         max_mapnr = last_valid_pfn;
1803 #endif
1804         kernel_physical_mapping_init();
1805
1806         {
1807                 unsigned long max_zone_pfns[MAX_NR_ZONES];
1808
1809                 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
1810
1811                 max_zone_pfns[ZONE_NORMAL] = end_pfn;
1812
1813                 free_area_init_nodes(max_zone_pfns);
1814         }
1815
1816         printk("Booting Linux...\n");
1817
1818         central_probe();
1819         cpu_probe();
1820 }
1821
1822 int __init page_in_phys_avail(unsigned long paddr)
1823 {
1824         int i;
1825
1826         paddr &= PAGE_MASK;
1827
1828         for (i = 0; i < pavail_ents; i++) {
1829                 unsigned long start, end;
1830
1831                 start = pavail[i].phys_addr;
1832                 end = start + pavail[i].reg_size;
1833
1834                 if (paddr >= start && paddr < end)
1835                         return 1;
1836         }
1837         if (paddr >= kern_base && paddr < (kern_base + kern_size))
1838                 return 1;
1839 #ifdef CONFIG_BLK_DEV_INITRD
1840         if (paddr >= __pa(initrd_start) &&
1841             paddr < __pa(PAGE_ALIGN(initrd_end)))
1842                 return 1;
1843 #endif
1844
1845         return 0;
1846 }
1847
1848 static struct linux_prom64_registers pavail_rescan[MAX_BANKS] __initdata;
1849 static int pavail_rescan_ents __initdata;
1850
1851 /* Certain OBP calls, such as fetching "available" properties, can
1852  * claim physical memory.  So, along with initializing the valid
1853  * address bitmap, what we do here is refetch the physical available
1854  * memory list again, and make sure it provides at least as much
1855  * memory as 'pavail' does.
1856  */
1857 static void setup_valid_addr_bitmap_from_pavail(void)
1858 {
1859         int i;
1860
1861         read_obp_memory("available", &pavail_rescan[0], &pavail_rescan_ents);
1862
1863         for (i = 0; i < pavail_ents; i++) {
1864                 unsigned long old_start, old_end;
1865
1866                 old_start = pavail[i].phys_addr;
1867                 old_end = old_start + pavail[i].reg_size;
1868                 while (old_start < old_end) {
1869                         int n;
1870
1871                         for (n = 0; n < pavail_rescan_ents; n++) {
1872                                 unsigned long new_start, new_end;
1873
1874                                 new_start = pavail_rescan[n].phys_addr;
1875                                 new_end = new_start +
1876                                         pavail_rescan[n].reg_size;
1877
1878                                 if (new_start <= old_start &&
1879                                     new_end >= (old_start + PAGE_SIZE)) {
1880                                         set_bit(old_start >> 22,
1881                                                 sparc64_valid_addr_bitmap);
1882                                         goto do_next_page;
1883                                 }
1884                         }
1885
1886                         prom_printf("mem_init: Lost memory in pavail\n");
1887                         prom_printf("mem_init: OLD start[%lx] size[%lx]\n",
1888                                     pavail[i].phys_addr,
1889                                     pavail[i].reg_size);
1890                         prom_printf("mem_init: NEW start[%lx] size[%lx]\n",
1891                                     pavail_rescan[i].phys_addr,
1892                                     pavail_rescan[i].reg_size);
1893                         prom_printf("mem_init: Cannot continue, aborting.\n");
1894                         prom_halt();
1895
1896                 do_next_page:
1897                         old_start += PAGE_SIZE;
1898                 }
1899         }
1900 }
1901
1902 void __init mem_init(void)
1903 {
1904         unsigned long codepages, datapages, initpages;
1905         unsigned long addr, last;
1906         int i;
1907
1908         i = last_valid_pfn >> ((22 - PAGE_SHIFT) + 6);
1909         i += 1;
1910         sparc64_valid_addr_bitmap = (unsigned long *) alloc_bootmem(i << 3);
1911         if (sparc64_valid_addr_bitmap == NULL) {
1912                 prom_printf("mem_init: Cannot alloc valid_addr_bitmap.\n");
1913                 prom_halt();
1914         }
1915         memset(sparc64_valid_addr_bitmap, 0, i << 3);
1916
1917         addr = PAGE_OFFSET + kern_base;
1918         last = PAGE_ALIGN(kern_size) + addr;
1919         while (addr < last) {
1920                 set_bit(__pa(addr) >> 22, sparc64_valid_addr_bitmap);
1921                 addr += PAGE_SIZE;
1922         }
1923
1924         setup_valid_addr_bitmap_from_pavail();
1925
1926         high_memory = __va(last_valid_pfn << PAGE_SHIFT);
1927
1928 #ifdef CONFIG_NEED_MULTIPLE_NODES
1929         for_each_online_node(i) {
1930                 if (NODE_DATA(i)->node_spanned_pages != 0) {
1931                         totalram_pages +=
1932                                 free_all_bootmem_node(NODE_DATA(i));
1933                 }
1934         }
1935 #else
1936         totalram_pages = free_all_bootmem();
1937 #endif
1938
1939         /* We subtract one to account for the mem_map_zero page
1940          * allocated below.
1941          */
1942         totalram_pages -= 1;
1943         num_physpages = totalram_pages;
1944
1945         /*
1946          * Set up the zero page, mark it reserved, so that page count
1947          * is not manipulated when freeing the page from user ptes.
1948          */
1949         mem_map_zero = alloc_pages(GFP_KERNEL|__GFP_ZERO, 0);
1950         if (mem_map_zero == NULL) {
1951                 prom_printf("paging_init: Cannot alloc zero page.\n");
1952                 prom_halt();
1953         }
1954         SetPageReserved(mem_map_zero);
1955
1956         codepages = (((unsigned long) _etext) - ((unsigned long) _start));
1957         codepages = PAGE_ALIGN(codepages) >> PAGE_SHIFT;
1958         datapages = (((unsigned long) _edata) - ((unsigned long) _etext));
1959         datapages = PAGE_ALIGN(datapages) >> PAGE_SHIFT;
1960         initpages = (((unsigned long) __init_end) - ((unsigned long) __init_begin));
1961         initpages = PAGE_ALIGN(initpages) >> PAGE_SHIFT;
1962
1963         printk("Memory: %luk available (%ldk kernel code, %ldk data, %ldk init) [%016lx,%016lx]\n",
1964                nr_free_pages() << (PAGE_SHIFT-10),
1965                codepages << (PAGE_SHIFT-10),
1966                datapages << (PAGE_SHIFT-10),
1967                initpages << (PAGE_SHIFT-10),
1968                PAGE_OFFSET, (last_valid_pfn << PAGE_SHIFT));
1969
1970         if (tlb_type == cheetah || tlb_type == cheetah_plus)
1971                 cheetah_ecache_flush_init();
1972 }
1973
1974 void free_initmem(void)
1975 {
1976         unsigned long addr, initend;
1977
1978         /*
1979          * The init section is aligned to 8k in vmlinux.lds. Page align for >8k pagesizes.
1980          */
1981         addr = PAGE_ALIGN((unsigned long)(__init_begin));
1982         initend = (unsigned long)(__init_end) & PAGE_MASK;
1983         for (; addr < initend; addr += PAGE_SIZE) {
1984                 unsigned long page;
1985                 struct page *p;
1986
1987                 page = (addr +
1988                         ((unsigned long) __va(kern_base)) -
1989                         ((unsigned long) KERNBASE));
1990                 memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
1991                 p = virt_to_page(page);
1992
1993                 ClearPageReserved(p);
1994                 init_page_count(p);
1995                 __free_page(p);
1996                 num_physpages++;
1997                 totalram_pages++;
1998         }
1999 }
2000
2001 #ifdef CONFIG_BLK_DEV_INITRD
2002 void free_initrd_mem(unsigned long start, unsigned long end)
2003 {
2004         if (start < end)
2005                 printk ("Freeing initrd memory: %ldk freed\n", (end - start) >> 10);
2006         for (; start < end; start += PAGE_SIZE) {
2007                 struct page *p = virt_to_page(start);
2008
2009                 ClearPageReserved(p);
2010                 init_page_count(p);
2011                 __free_page(p);
2012                 num_physpages++;
2013                 totalram_pages++;
2014         }
2015 }
2016 #endif
2017
2018 #define _PAGE_CACHE_4U  (_PAGE_CP_4U | _PAGE_CV_4U)
2019 #define _PAGE_CACHE_4V  (_PAGE_CP_4V | _PAGE_CV_4V)
2020 #define __DIRTY_BITS_4U  (_PAGE_MODIFIED_4U | _PAGE_WRITE_4U | _PAGE_W_4U)
2021 #define __DIRTY_BITS_4V  (_PAGE_MODIFIED_4V | _PAGE_WRITE_4V | _PAGE_W_4V)
2022 #define __ACCESS_BITS_4U (_PAGE_ACCESSED_4U | _PAGE_READ_4U | _PAGE_R)
2023 #define __ACCESS_BITS_4V (_PAGE_ACCESSED_4V | _PAGE_READ_4V | _PAGE_R)
2024
2025 pgprot_t PAGE_KERNEL __read_mostly;
2026 EXPORT_SYMBOL(PAGE_KERNEL);
2027
2028 pgprot_t PAGE_KERNEL_LOCKED __read_mostly;
2029 pgprot_t PAGE_COPY __read_mostly;
2030
2031 pgprot_t PAGE_SHARED __read_mostly;
2032 EXPORT_SYMBOL(PAGE_SHARED);
2033
2034 pgprot_t PAGE_EXEC __read_mostly;
2035 unsigned long pg_iobits __read_mostly;
2036
2037 unsigned long _PAGE_IE __read_mostly;
2038 EXPORT_SYMBOL(_PAGE_IE);
2039
2040 unsigned long _PAGE_E __read_mostly;
2041 EXPORT_SYMBOL(_PAGE_E);
2042
2043 unsigned long _PAGE_CACHE __read_mostly;
2044 EXPORT_SYMBOL(_PAGE_CACHE);
2045
2046 #ifdef CONFIG_SPARSEMEM_VMEMMAP
2047
2048 #define VMEMMAP_CHUNK_SHIFT     22
2049 #define VMEMMAP_CHUNK           (1UL << VMEMMAP_CHUNK_SHIFT)
2050 #define VMEMMAP_CHUNK_MASK      ~(VMEMMAP_CHUNK - 1UL)
2051 #define VMEMMAP_ALIGN(x)        (((x)+VMEMMAP_CHUNK-1UL)&VMEMMAP_CHUNK_MASK)
2052
2053 #define VMEMMAP_SIZE    ((((1UL << MAX_PHYSADDR_BITS) >> PAGE_SHIFT) * \
2054                           sizeof(struct page *)) >> VMEMMAP_CHUNK_SHIFT)
2055 unsigned long vmemmap_table[VMEMMAP_SIZE];
2056
2057 int __meminit vmemmap_populate(struct page *start, unsigned long nr, int node)
2058 {
2059         unsigned long vstart = (unsigned long) start;
2060         unsigned long vend = (unsigned long) (start + nr);
2061         unsigned long phys_start = (vstart - VMEMMAP_BASE);
2062         unsigned long phys_end = (vend - VMEMMAP_BASE);
2063         unsigned long addr = phys_start & VMEMMAP_CHUNK_MASK;
2064         unsigned long end = VMEMMAP_ALIGN(phys_end);
2065         unsigned long pte_base;
2066
2067         pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2068                     _PAGE_CP_4U | _PAGE_CV_4U |
2069                     _PAGE_P_4U | _PAGE_W_4U);
2070         if (tlb_type == hypervisor)
2071                 pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2072                             _PAGE_CP_4V | _PAGE_CV_4V |
2073                             _PAGE_P_4V | _PAGE_W_4V);
2074
2075         for (; addr < end; addr += VMEMMAP_CHUNK) {
2076                 unsigned long *vmem_pp =
2077                         vmemmap_table + (addr >> VMEMMAP_CHUNK_SHIFT);
2078                 void *block;
2079
2080                 if (!(*vmem_pp & _PAGE_VALID)) {
2081                         block = vmemmap_alloc_block(1UL << 22, node);
2082                         if (!block)
2083                                 return -ENOMEM;
2084
2085                         *vmem_pp = pte_base | __pa(block);
2086
2087                         printk(KERN_INFO "[%p-%p] page_structs=%lu "
2088                                "node=%d entry=%lu/%lu\n", start, block, nr,
2089                                node,
2090                                addr >> VMEMMAP_CHUNK_SHIFT,
2091                                VMEMMAP_SIZE >> VMEMMAP_CHUNK_SHIFT);
2092                 }
2093         }
2094         return 0;
2095 }
2096 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
2097
2098 static void prot_init_common(unsigned long page_none,
2099                              unsigned long page_shared,
2100                              unsigned long page_copy,
2101                              unsigned long page_readonly,
2102                              unsigned long page_exec_bit)
2103 {
2104         PAGE_COPY = __pgprot(page_copy);
2105         PAGE_SHARED = __pgprot(page_shared);
2106
2107         protection_map[0x0] = __pgprot(page_none);
2108         protection_map[0x1] = __pgprot(page_readonly & ~page_exec_bit);
2109         protection_map[0x2] = __pgprot(page_copy & ~page_exec_bit);
2110         protection_map[0x3] = __pgprot(page_copy & ~page_exec_bit);
2111         protection_map[0x4] = __pgprot(page_readonly);
2112         protection_map[0x5] = __pgprot(page_readonly);
2113         protection_map[0x6] = __pgprot(page_copy);
2114         protection_map[0x7] = __pgprot(page_copy);
2115         protection_map[0x8] = __pgprot(page_none);
2116         protection_map[0x9] = __pgprot(page_readonly & ~page_exec_bit);
2117         protection_map[0xa] = __pgprot(page_shared & ~page_exec_bit);
2118         protection_map[0xb] = __pgprot(page_shared & ~page_exec_bit);
2119         protection_map[0xc] = __pgprot(page_readonly);
2120         protection_map[0xd] = __pgprot(page_readonly);
2121         protection_map[0xe] = __pgprot(page_shared);
2122         protection_map[0xf] = __pgprot(page_shared);
2123 }
2124
2125 static void __init sun4u_pgprot_init(void)
2126 {
2127         unsigned long page_none, page_shared, page_copy, page_readonly;
2128         unsigned long page_exec_bit;
2129
2130         PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2131                                 _PAGE_CACHE_4U | _PAGE_P_4U |
2132                                 __ACCESS_BITS_4U | __DIRTY_BITS_4U |
2133                                 _PAGE_EXEC_4U);
2134         PAGE_KERNEL_LOCKED = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2135                                        _PAGE_CACHE_4U | _PAGE_P_4U |
2136                                        __ACCESS_BITS_4U | __DIRTY_BITS_4U |
2137                                        _PAGE_EXEC_4U | _PAGE_L_4U);
2138         PAGE_EXEC = __pgprot(_PAGE_EXEC_4U);
2139
2140         _PAGE_IE = _PAGE_IE_4U;
2141         _PAGE_E = _PAGE_E_4U;
2142         _PAGE_CACHE = _PAGE_CACHE_4U;
2143
2144         pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4U | __DIRTY_BITS_4U |
2145                      __ACCESS_BITS_4U | _PAGE_E_4U);
2146
2147 #ifdef CONFIG_DEBUG_PAGEALLOC
2148         kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZBITS_4U) ^
2149                 0xfffff80000000000;
2150 #else
2151         kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4U) ^
2152                 0xfffff80000000000;
2153 #endif
2154         kern_linear_pte_xor[0] |= (_PAGE_CP_4U | _PAGE_CV_4U |
2155                                    _PAGE_P_4U | _PAGE_W_4U);
2156
2157         /* XXX Should use 256MB on Panther. XXX */
2158         kern_linear_pte_xor[1] = kern_linear_pte_xor[0];
2159
2160         _PAGE_SZBITS = _PAGE_SZBITS_4U;
2161         _PAGE_ALL_SZ_BITS =  (_PAGE_SZ4MB_4U | _PAGE_SZ512K_4U |
2162                               _PAGE_SZ64K_4U | _PAGE_SZ8K_4U |
2163                               _PAGE_SZ32MB_4U | _PAGE_SZ256MB_4U);
2164
2165
2166         page_none = _PAGE_PRESENT_4U | _PAGE_ACCESSED_4U | _PAGE_CACHE_4U;
2167         page_shared = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2168                        __ACCESS_BITS_4U | _PAGE_WRITE_4U | _PAGE_EXEC_4U);
2169         page_copy   = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2170                        __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2171         page_readonly   = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2172                            __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2173
2174         page_exec_bit = _PAGE_EXEC_4U;
2175
2176         prot_init_common(page_none, page_shared, page_copy, page_readonly,
2177                          page_exec_bit);
2178 }
2179
2180 static void __init sun4v_pgprot_init(void)
2181 {
2182         unsigned long page_none, page_shared, page_copy, page_readonly;
2183         unsigned long page_exec_bit;
2184
2185         PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4V | _PAGE_VALID |
2186                                 _PAGE_CACHE_4V | _PAGE_P_4V |
2187                                 __ACCESS_BITS_4V | __DIRTY_BITS_4V |
2188                                 _PAGE_EXEC_4V);
2189         PAGE_KERNEL_LOCKED = PAGE_KERNEL;
2190         PAGE_EXEC = __pgprot(_PAGE_EXEC_4V);
2191
2192         _PAGE_IE = _PAGE_IE_4V;
2193         _PAGE_E = _PAGE_E_4V;
2194         _PAGE_CACHE = _PAGE_CACHE_4V;
2195
2196 #ifdef CONFIG_DEBUG_PAGEALLOC
2197         kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZBITS_4V) ^
2198                 0xfffff80000000000;
2199 #else
2200         kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4V) ^
2201                 0xfffff80000000000;
2202 #endif
2203         kern_linear_pte_xor[0] |= (_PAGE_CP_4V | _PAGE_CV_4V |
2204                                    _PAGE_P_4V | _PAGE_W_4V);
2205
2206 #ifdef CONFIG_DEBUG_PAGEALLOC
2207         kern_linear_pte_xor[1] = (_PAGE_VALID | _PAGE_SZBITS_4V) ^
2208                 0xfffff80000000000;
2209 #else
2210         kern_linear_pte_xor[1] = (_PAGE_VALID | _PAGE_SZ256MB_4V) ^
2211                 0xfffff80000000000;
2212 #endif
2213         kern_linear_pte_xor[1] |= (_PAGE_CP_4V | _PAGE_CV_4V |
2214                                    _PAGE_P_4V | _PAGE_W_4V);
2215
2216         pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4V | __DIRTY_BITS_4V |
2217                      __ACCESS_BITS_4V | _PAGE_E_4V);
2218
2219         _PAGE_SZBITS = _PAGE_SZBITS_4V;
2220         _PAGE_ALL_SZ_BITS = (_PAGE_SZ16GB_4V | _PAGE_SZ2GB_4V |
2221                              _PAGE_SZ256MB_4V | _PAGE_SZ32MB_4V |
2222                              _PAGE_SZ4MB_4V | _PAGE_SZ512K_4V |
2223                              _PAGE_SZ64K_4V | _PAGE_SZ8K_4V);
2224
2225         page_none = _PAGE_PRESENT_4V | _PAGE_ACCESSED_4V | _PAGE_CACHE_4V;
2226         page_shared = (_PAGE_VALID | _PAGE_PRESENT_4V | _PAGE_CACHE_4V |
2227                        __ACCESS_BITS_4V | _PAGE_WRITE_4V | _PAGE_EXEC_4V);
2228         page_copy   = (_PAGE_VALID | _PAGE_PRESENT_4V | _PAGE_CACHE_4V |
2229                        __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2230         page_readonly = (_PAGE_VALID | _PAGE_PRESENT_4V | _PAGE_CACHE_4V |
2231                          __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2232
2233         page_exec_bit = _PAGE_EXEC_4V;
2234
2235         prot_init_common(page_none, page_shared, page_copy, page_readonly,
2236                          page_exec_bit);
2237 }
2238
2239 unsigned long pte_sz_bits(unsigned long sz)
2240 {
2241         if (tlb_type == hypervisor) {
2242                 switch (sz) {
2243                 case 8 * 1024:
2244                 default:
2245                         return _PAGE_SZ8K_4V;
2246                 case 64 * 1024:
2247                         return _PAGE_SZ64K_4V;
2248                 case 512 * 1024:
2249                         return _PAGE_SZ512K_4V;
2250                 case 4 * 1024 * 1024:
2251                         return _PAGE_SZ4MB_4V;
2252                 };
2253         } else {
2254                 switch (sz) {
2255                 case 8 * 1024:
2256                 default:
2257                         return _PAGE_SZ8K_4U;
2258                 case 64 * 1024:
2259                         return _PAGE_SZ64K_4U;
2260                 case 512 * 1024:
2261                         return _PAGE_SZ512K_4U;
2262                 case 4 * 1024 * 1024:
2263                         return _PAGE_SZ4MB_4U;
2264                 };
2265         }
2266 }
2267
2268 pte_t mk_pte_io(unsigned long page, pgprot_t prot, int space, unsigned long page_size)
2269 {
2270         pte_t pte;
2271
2272         pte_val(pte)  = page | pgprot_val(pgprot_noncached(prot));
2273         pte_val(pte) |= (((unsigned long)space) << 32);
2274         pte_val(pte) |= pte_sz_bits(page_size);
2275
2276         return pte;
2277 }
2278
2279 static unsigned long kern_large_tte(unsigned long paddr)
2280 {
2281         unsigned long val;
2282
2283         val = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2284                _PAGE_CP_4U | _PAGE_CV_4U | _PAGE_P_4U |
2285                _PAGE_EXEC_4U | _PAGE_L_4U | _PAGE_W_4U);
2286         if (tlb_type == hypervisor)
2287                 val = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2288                        _PAGE_CP_4V | _PAGE_CV_4V | _PAGE_P_4V |
2289                        _PAGE_EXEC_4V | _PAGE_W_4V);
2290
2291         return val | paddr;
2292 }
2293
2294 /* If not locked, zap it. */
2295 void __flush_tlb_all(void)
2296 {
2297         unsigned long pstate;
2298         int i;
2299
2300         __asm__ __volatile__("flushw\n\t"
2301                              "rdpr      %%pstate, %0\n\t"
2302                              "wrpr      %0, %1, %%pstate"
2303                              : "=r" (pstate)
2304                              : "i" (PSTATE_IE));
2305         if (tlb_type == hypervisor) {
2306                 sun4v_mmu_demap_all();
2307         } else if (tlb_type == spitfire) {
2308                 for (i = 0; i < 64; i++) {
2309                         /* Spitfire Errata #32 workaround */
2310                         /* NOTE: Always runs on spitfire, so no
2311                          *       cheetah+ page size encodings.
2312                          */
2313                         __asm__ __volatile__("stxa      %0, [%1] %2\n\t"
2314                                              "flush     %%g6"
2315                                              : /* No outputs */
2316                                              : "r" (0),
2317                                              "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2318
2319                         if (!(spitfire_get_dtlb_data(i) & _PAGE_L_4U)) {
2320                                 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2321                                                      "membar #Sync"
2322                                                      : /* no outputs */
2323                                                      : "r" (TLB_TAG_ACCESS), "i" (ASI_DMMU));
2324                                 spitfire_put_dtlb_data(i, 0x0UL);
2325                         }
2326
2327                         /* Spitfire Errata #32 workaround */
2328                         /* NOTE: Always runs on spitfire, so no
2329                          *       cheetah+ page size encodings.
2330                          */
2331                         __asm__ __volatile__("stxa      %0, [%1] %2\n\t"
2332                                              "flush     %%g6"
2333                                              : /* No outputs */
2334                                              : "r" (0),
2335                                              "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2336
2337                         if (!(spitfire_get_itlb_data(i) & _PAGE_L_4U)) {
2338                                 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2339                                                      "membar #Sync"
2340                                                      : /* no outputs */
2341                                                      : "r" (TLB_TAG_ACCESS), "i" (ASI_IMMU));
2342                                 spitfire_put_itlb_data(i, 0x0UL);
2343                         }
2344                 }
2345         } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
2346                 cheetah_flush_dtlb_all();
2347                 cheetah_flush_itlb_all();
2348         }
2349         __asm__ __volatile__("wrpr      %0, 0, %%pstate"
2350                              : : "r" (pstate));
2351 }
2352
2353 #ifdef CONFIG_MEMORY_HOTPLUG
2354
2355 void online_page(struct page *page)
2356 {
2357         ClearPageReserved(page);
2358         init_page_count(page);
2359         __free_page(page);
2360         totalram_pages++;
2361         num_physpages++;
2362 }
2363
2364 #endif /* CONFIG_MEMORY_HOTPLUG */