arch/sparc64/mm/init.c

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