arch/sparc/mm/init_64.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/initrd.h>
  17 #include <linux/swap.h>
  18 #include <linux/pagemap.h>
  19 #include <linux/poison.h>
  20 #include <linux/fs.h>
  21 #include <linux/seq_file.h>
  22 #include <linux/kprobes.h>
  23 #include <linux/cache.h>
  24 #include <linux/sort.h>
  25 #include <linux/percpu.h>
  26 #include <linux/memblock.h>
  27 #include <linux/mmzone.h>
  28 #include <linux/gfp.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/mdesc.h>
  50 #include <asm/cpudata.h>
  51 #include <asm/irq.h>
  52
  53 #include "init_64.h"
  54
  55 unsigned long kern_linear_pte_xor[2] __read_mostly;
  56
  57 /* A bitmap, one bit for every 256MB of physical memory.  If the bit
  58  * is clear, we should use a 4MB page (via kern_linear_pte_xor[0]) else
  59  * if set we should use a 256MB page (via kern_linear_pte_xor[1]).
  60  */
  61 unsigned long kpte_linear_bitmap[KPTE_BITMAP_BYTES / sizeof(unsigned long)];
  62
  63 #ifndef CONFIG_DEBUG_PAGEALLOC
  64 /* A special kernel TSB for 4MB and 256MB linear mappings.
  65  * Space is allocated for this right after the trap table
  66  * in arch/sparc64/kernel/head.S
  67  */
  68 extern struct tsb swapper_4m_tsb[KERNEL_TSB4M_NENTRIES];
  69 #endif
  70
  71 #define MAX_BANKS       32
  72
  73 static struct linux_prom64_registers pavail[MAX_BANKS] __devinitdata;
  74 static int pavail_ents __devinitdata;
  75
  76 static int cmp_p64(const void *a, const void *b)
  77 {
  78         const struct linux_prom64_registers *x = a, *y = b;
  79
  80         if (x->phys_addr > y->phys_addr)
  81                 return 1;
  82         if (x->phys_addr < y->phys_addr)
  83                 return -1;
  84         return 0;
  85 }
  86
  87 static void __init read_obp_memory(const char *property,
  88                                    struct linux_prom64_registers *regs,
  89                                    int *num_ents)
  90 {
  91         phandle node = prom_finddevice("/memory");
  92         int prop_size = prom_getproplen(node, property);
  93         int ents, ret, i;
  94
  95         ents = prop_size / sizeof(struct linux_prom64_registers);
  96         if (ents > MAX_BANKS) {
  97                 prom_printf("The machine has more %s property entries than "
  98                             "this kernel can support (%d).\n",
  99                             property, MAX_BANKS);
 100                 prom_halt();
 101         }
 102
 103         ret = prom_getproperty(node, property, (char *) regs, prop_size);
 104         if (ret == -1) {
 105                 prom_printf("Couldn't get %s property from /memory.\n");
 106                 prom_halt();
 107         }
 108
 109         /* Sanitize what we got from the firmware, by page aligning
 110          * everything.
 111          */
 112         for (i = 0; i < ents; i++) {
 113                 unsigned long base, size;
 114
 115                 base = regs[i].phys_addr;
 116                 size = regs[i].reg_size;
 117
 118                 size &= PAGE_MASK;
 119                 if (base & ~PAGE_MASK) {
 120                         unsigned long new_base = PAGE_ALIGN(base);
 121
 122                         size -= new_base - base;
 123                         if ((long) size < 0L)
 124                                 size = 0UL;
 125                         base = new_base;
 126                 }
 127                 if (size == 0UL) {
 128                         /* If it is empty, simply get rid of it.
 129                          * This simplifies the logic of the other
 130                          * functions that process these arrays.
 131                          */
 132                         memmove(&regs[i], &regs[i + 1],
 133                                 (ents - i - 1) * sizeof(regs[0]));
 134                         i--;
 135                         ents--;
 136                         continue;
 137                 }
 138                 regs[i].phys_addr = base;
 139                 regs[i].reg_size = size;
 140         }
 141
 142         *num_ents = ents;
 143
 144         sort(regs, ents, sizeof(struct linux_prom64_registers),
 145              cmp_p64, NULL);
 146 }
 147
 148 unsigned long sparc64_valid_addr_bitmap[VALID_ADDR_BITMAP_BYTES /
 149                                         sizeof(unsigned long)];
 150 EXPORT_SYMBOL(sparc64_valid_addr_bitmap);
 151
 152 /* Kernel physical address base and size in bytes.  */
 153 unsigned long kern_base __read_mostly;
 154 unsigned long kern_size __read_mostly;
 155
 156 /* Initial ramdisk setup */
 157 extern unsigned long sparc_ramdisk_image64;
 158 extern unsigned int sparc_ramdisk_image;
 159 extern unsigned int sparc_ramdisk_size;
 160
 161 struct page *mem_map_zero __read_mostly;
 162 EXPORT_SYMBOL(mem_map_zero);
 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                              "bne,pn    %%xcc, 1b\n\t"
 220                              " nop"
 221                              : /* no outputs */
 222                              : "r" (mask), "r" (non_cpu_bits), "r" (&page->flags)
 223                              : "g1", "g7");
 224 }
 225
 226 static inline void clear_dcache_dirty_cpu(struct page *page, unsigned long cpu)
 227 {
 228         unsigned long mask = (1UL << PG_dcache_dirty);
 229
 230         __asm__ __volatile__("! test_and_clear_dcache_dirty\n"
 231                              "1:\n\t"
 232                              "ldx       [%2], %%g7\n\t"
 233                              "srlx      %%g7, %4, %%g1\n\t"
 234                              "and       %%g1, %3, %%g1\n\t"
 235                              "cmp       %%g1, %0\n\t"
 236                              "bne,pn    %%icc, 2f\n\t"
 237                              " andn     %%g7, %1, %%g1\n\t"
 238                              "casx      [%2], %%g7, %%g1\n\t"
 239                              "cmp       %%g7, %%g1\n\t"
 240                              "bne,pn    %%xcc, 1b\n\t"
 241                              " nop\n"
 242                              "2:"
 243                              : /* no outputs */
 244                              : "r" (cpu), "r" (mask), "r" (&page->flags),
 245                                "i" (PG_dcache_cpu_mask),
 246                                "i" (PG_dcache_cpu_shift)
 247                              : "g1", "g7");
 248 }
 249
 250 static inline void tsb_insert(struct tsb *ent, unsigned long tag, unsigned long pte)
 251 {
 252         unsigned long tsb_addr = (unsigned long) ent;
 253
 254         if (tlb_type == cheetah_plus || tlb_type == hypervisor)
 255                 tsb_addr = __pa(tsb_addr);
 256
 257         __tsb_insert(tsb_addr, tag, pte);
 258 }
 259
 260 unsigned long _PAGE_ALL_SZ_BITS __read_mostly;
 261 unsigned long _PAGE_SZBITS __read_mostly;
 262
 263 static void flush_dcache(unsigned long pfn)
 264 {
 265         struct page *page;
 266
 267         page = pfn_to_page(pfn);
 268         if (page) {
 269                 unsigned long pg_flags;
 270
 271                 pg_flags = page->flags;
 272                 if (pg_flags & (1UL << PG_dcache_dirty)) {
 273                         int cpu = ((pg_flags >> PG_dcache_cpu_shift) &
 274                                    PG_dcache_cpu_mask);
 275                         int this_cpu = get_cpu();
 276
 277                         /* This is just to optimize away some function calls
 278                          * in the SMP case.
 279                          */
 280                         if (cpu == this_cpu)
 281                                 flush_dcache_page_impl(page);
 282                         else
 283                                 smp_flush_dcache_page_impl(page, cpu);
 284
 285                         clear_dcache_dirty_cpu(page, cpu);
 286
 287                         put_cpu();
 288                 }
 289         }
 290 }
 291
 292 void update_mmu_cache(struct vm_area_struct *vma, unsigned long address, pte_t *ptep)
 293 {
 294         struct mm_struct *mm;
 295         struct tsb *tsb;
 296         unsigned long tag, flags;
 297         unsigned long tsb_index, tsb_hash_shift;
 298         pte_t pte = *ptep;
 299
 300         if (tlb_type != hypervisor) {
 301                 unsigned long pfn = pte_pfn(pte);
 302
 303                 if (pfn_valid(pfn))
 304                         flush_dcache(pfn);
 305         }
 306
 307         mm = vma->vm_mm;
 308
 309         tsb_index = MM_TSB_BASE;
 310         tsb_hash_shift = PAGE_SHIFT;
 311
 312         spin_lock_irqsave(&mm->context.lock, flags);
 313
 314 #ifdef CONFIG_HUGETLB_PAGE
 315         if (mm->context.tsb_block[MM_TSB_HUGE].tsb != NULL) {
 316                 if ((tlb_type == hypervisor &&
 317                      (pte_val(pte) & _PAGE_SZALL_4V) == _PAGE_SZHUGE_4V) ||
 318                     (tlb_type != hypervisor &&
 319                      (pte_val(pte) & _PAGE_SZALL_4U) == _PAGE_SZHUGE_4U)) {
 320                         tsb_index = MM_TSB_HUGE;
 321                         tsb_hash_shift = HPAGE_SHIFT;
 322                 }
 323         }
 324 #endif
 325
 326         tsb = mm->context.tsb_block[tsb_index].tsb;
 327         tsb += ((address >> tsb_hash_shift) &
 328                 (mm->context.tsb_block[tsb_index].tsb_nentries - 1UL));
 329         tag = (address >> 22UL);
 330         tsb_insert(tsb, tag, pte_val(pte));
 331
 332         spin_unlock_irqrestore(&mm->context.lock, flags);
 333 }
 334
 335 void flush_dcache_page(struct page *page)
 336 {
 337         struct address_space *mapping;
 338         int this_cpu;
 339
 340         if (tlb_type == hypervisor)
 341                 return;
 342
 343         /* Do not bother with the expensive D-cache flush if it
 344          * is merely the zero page.  The 'bigcore' testcase in GDB
 345          * causes this case to run millions of times.
 346          */
 347         if (page == ZERO_PAGE(0))
 348                 return;
 349
 350         this_cpu = get_cpu();
 351
 352         mapping = page_mapping(page);
 353         if (mapping && !mapping_mapped(mapping)) {
 354                 int dirty = test_bit(PG_dcache_dirty, &page->flags);
 355                 if (dirty) {
 356                         int dirty_cpu = dcache_dirty_cpu(page);
 357
 358                         if (dirty_cpu == this_cpu)
 359                                 goto out;
 360                         smp_flush_dcache_page_impl(page, dirty_cpu);
 361                 }
 362                 set_dcache_dirty(page, this_cpu);
 363         } else {
 364                 /* We could delay the flush for the !page_mapping
 365                  * case too.  But that case is for exec env/arg
 366                  * pages and those are %99 certainly going to get
 367                  * faulted into the tlb (and thus flushed) anyways.
 368                  */
 369                 flush_dcache_page_impl(page);
 370         }
 371
 372 out:
 373         put_cpu();
 374 }
 375 EXPORT_SYMBOL(flush_dcache_page);
 376
 377 void __kprobes flush_icache_range(unsigned long start, unsigned long end)
 378 {
 379         /* Cheetah and Hypervisor platform cpus have coherent I-cache. */
 380         if (tlb_type == spitfire) {
 381                 unsigned long kaddr;
 382
 383                 /* This code only runs on Spitfire cpus so this is
 384                  * why we can assume _PAGE_PADDR_4U.
 385                  */
 386                 for (kaddr = start; kaddr < end; kaddr += PAGE_SIZE) {
 387                         unsigned long paddr, mask = _PAGE_PADDR_4U;
 388
 389                         if (kaddr >= PAGE_OFFSET)
 390                                 paddr = kaddr & mask;
 391                         else {
 392                                 pgd_t *pgdp = pgd_offset_k(kaddr);
 393                                 pud_t *pudp = pud_offset(pgdp, kaddr);
 394                                 pmd_t *pmdp = pmd_offset(pudp, kaddr);
 395                                 pte_t *ptep = pte_offset_kernel(pmdp, kaddr);
 396
 397                                 paddr = pte_val(*ptep) & mask;
 398                         }
 399                         __flush_icache_page(paddr);
 400                 }
 401         }
 402 }
 403 EXPORT_SYMBOL(flush_icache_range);
 404
 405 void mmu_info(struct seq_file *m)
 406 {
 407         if (tlb_type == cheetah)
 408                 seq_printf(m, "MMU Type\t: Cheetah\n");
 409         else if (tlb_type == cheetah_plus)
 410                 seq_printf(m, "MMU Type\t: Cheetah+\n");
 411         else if (tlb_type == spitfire)
 412                 seq_printf(m, "MMU Type\t: Spitfire\n");
 413         else if (tlb_type == hypervisor)
 414                 seq_printf(m, "MMU Type\t: Hypervisor (sun4v)\n");
 415         else
 416                 seq_printf(m, "MMU Type\t: ???\n");
 417
 418 #ifdef CONFIG_DEBUG_DCFLUSH
 419         seq_printf(m, "DCPageFlushes\t: %d\n",
 420                    atomic_read(&dcpage_flushes));
 421 #ifdef CONFIG_SMP
 422         seq_printf(m, "DCPageFlushesXC\t: %d\n",
 423                    atomic_read(&dcpage_flushes_xcall));
 424 #endif /* CONFIG_SMP */
 425 #endif /* CONFIG_DEBUG_DCFLUSH */
 426 }
 427
 428 struct linux_prom_translation prom_trans[512] __read_mostly;
 429 unsigned int prom_trans_ents __read_mostly;
 430
 431 unsigned long kern_locked_tte_data;
 432
 433 /* The obp translations are saved based on 8k pagesize, since obp can
 434  * use a mixture of pagesizes. Misses to the LOW_OBP_ADDRESS ->
 435  * HI_OBP_ADDRESS range are handled in ktlb.S.
 436  */
 437 static inline int in_obp_range(unsigned long vaddr)
 438 {
 439         return (vaddr >= LOW_OBP_ADDRESS &&
 440                 vaddr < HI_OBP_ADDRESS);
 441 }
 442
 443 static int cmp_ptrans(const void *a, const void *b)
 444 {
 445         const struct linux_prom_translation *x = a, *y = b;
 446
 447         if (x->virt > y->virt)
 448                 return 1;
 449         if (x->virt < y->virt)
 450                 return -1;
 451         return 0;
 452 }
 453
 454 /* Read OBP translations property into 'prom_trans[]'.  */
 455 static void __init read_obp_translations(void)
 456 {
 457         int n, node, ents, first, last, i;
 458
 459         node = prom_finddevice("/virtual-memory");
 460         n = prom_getproplen(node, "translations");
 461         if (unlikely(n == 0 || n == -1)) {
 462                 prom_printf("prom_mappings: Couldn't get size.\n");
 463                 prom_halt();
 464         }
 465         if (unlikely(n > sizeof(prom_trans))) {
 466                 prom_printf("prom_mappings: Size %Zd is too big.\n", n);
 467                 prom_halt();
 468         }
 469
 470         if ((n = prom_getproperty(node, "translations",
 471                                   (char *)&prom_trans[0],
 472                                   sizeof(prom_trans))) == -1) {
 473                 prom_printf("prom_mappings: Couldn't get property.\n");
 474                 prom_halt();
 475         }
 476
 477         n = n / sizeof(struct linux_prom_translation);
 478
 479         ents = n;
 480
 481         sort(prom_trans, ents, sizeof(struct linux_prom_translation),
 482              cmp_ptrans, NULL);
 483
 484         /* Now kick out all the non-OBP entries.  */
 485         for (i = 0; i < ents; i++) {
 486                 if (in_obp_range(prom_trans[i].virt))
 487                         break;
 488         }
 489         first = i;
 490         for (; i < ents; i++) {
 491                 if (!in_obp_range(prom_trans[i].virt))
 492                         break;
 493         }
 494         last = i;
 495
 496         for (i = 0; i < (last - first); i++) {
 497                 struct linux_prom_translation *src = &prom_trans[i + first];
 498                 struct linux_prom_translation *dest = &prom_trans[i];
 499
 500                 *dest = *src;
 501         }
 502         for (; i < ents; i++) {
 503                 struct linux_prom_translation *dest = &prom_trans[i];
 504                 dest->virt = dest->size = dest->data = 0x0UL;
 505         }
 506
 507         prom_trans_ents = last - first;
 508
 509         if (tlb_type == spitfire) {
 510                 /* Clear diag TTE bits. */
 511                 for (i = 0; i < prom_trans_ents; i++)
 512                         prom_trans[i].data &= ~0x0003fe0000000000UL;
 513         }
 514
 515         /* Force execute bit on.  */
 516         for (i = 0; i < prom_trans_ents; i++)
 517                 prom_trans[i].data |= (tlb_type == hypervisor ?
 518                                        _PAGE_EXEC_4V : _PAGE_EXEC_4U);
 519 }
 520
 521 static void __init hypervisor_tlb_lock(unsigned long vaddr,
 522                                        unsigned long pte,
 523                                        unsigned long mmu)
 524 {
 525         unsigned long ret = sun4v_mmu_map_perm_addr(vaddr, 0, pte, mmu);
 526
 527         if (ret != 0) {
 528                 prom_printf("hypervisor_tlb_lock[%lx:%lx:%lx:%lx]: "
 529                             "errors with %lx\n", vaddr, 0, pte, mmu, ret);
 530                 prom_halt();
 531         }
 532 }
 533
 534 static unsigned long kern_large_tte(unsigned long paddr);
 535
 536 static void __init remap_kernel(void)
 537 {
 538         unsigned long phys_page, tte_vaddr, tte_data;
 539         int i, tlb_ent = sparc64_highest_locked_tlbent();
 540
 541         tte_vaddr = (unsigned long) KERNBASE;
 542         phys_page = (prom_boot_mapping_phys_low >> 22UL) << 22UL;
 543         tte_data = kern_large_tte(phys_page);
 544
 545         kern_locked_tte_data = tte_data;
 546
 547         /* Now lock us into the TLBs via Hypervisor or OBP. */
 548         if (tlb_type == hypervisor) {
 549                 for (i = 0; i < num_kernel_image_mappings; i++) {
 550                         hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_DMMU);
 551                         hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_IMMU);
 552                         tte_vaddr += 0x400000;
 553                         tte_data += 0x400000;
 554                 }
 555         } else {
 556                 for (i = 0; i < num_kernel_image_mappings; i++) {
 557                         prom_dtlb_load(tlb_ent - i, tte_data, tte_vaddr);
 558                         prom_itlb_load(tlb_ent - i, tte_data, tte_vaddr);
 559                         tte_vaddr += 0x400000;
 560                         tte_data += 0x400000;
 561                 }
 562                 sparc64_highest_unlocked_tlb_ent = tlb_ent - i;
 563         }
 564         if (tlb_type == cheetah_plus) {
 565                 sparc64_kern_pri_context = (CTX_CHEETAH_PLUS_CTX0 |
 566                                             CTX_CHEETAH_PLUS_NUC);
 567                 sparc64_kern_pri_nuc_bits = CTX_CHEETAH_PLUS_NUC;
 568                 sparc64_kern_sec_context = CTX_CHEETAH_PLUS_CTX0;
 569         }
 570 }
 571
 572
 573 static void __init inherit_prom_mappings(void)
 574 {
 575         /* Now fixup OBP's idea about where we really are mapped. */
 576         printk("Remapping the kernel... ");
 577         remap_kernel();
 578         printk("done.\n");
 579 }
 580
 581 void prom_world(int enter)
 582 {
 583         if (!enter)
 584                 set_fs((mm_segment_t) { get_thread_current_ds() });
 585
 586         __asm__ __volatile__("flushw");
 587 }
 588
 589 void __flush_dcache_range(unsigned long start, unsigned long end)
 590 {
 591         unsigned long va;
 592
 593         if (tlb_type == spitfire) {
 594                 int n = 0;
 595
 596                 for (va = start; va < end; va += 32) {
 597                         spitfire_put_dcache_tag(va & 0x3fe0, 0x0);
 598                         if (++n >= 512)
 599                                 break;
 600                 }
 601         } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
 602                 start = __pa(start);
 603                 end = __pa(end);
 604                 for (va = start; va < end; va += 32)
 605                         __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
 606                                              "membar #Sync"
 607                                              : /* no outputs */
 608                                              : "r" (va),
 609                                                "i" (ASI_DCACHE_INVALIDATE));
 610         }
 611 }
 612 EXPORT_SYMBOL(__flush_dcache_range);
 613
 614 /* get_new_mmu_context() uses "cache + 1".  */
 615 DEFINE_SPINLOCK(ctx_alloc_lock);
 616 unsigned long tlb_context_cache = CTX_FIRST_VERSION - 1;
 617 #define MAX_CTX_NR      (1UL << CTX_NR_BITS)
 618 #define CTX_BMAP_SLOTS  BITS_TO_LONGS(MAX_CTX_NR)
 619 DECLARE_BITMAP(mmu_context_bmap, MAX_CTX_NR);
 620
 621 /* Caller does TLB context flushing on local CPU if necessary.
 622  * The caller also ensures that CTX_VALID(mm->context) is false.
 623  *
 624  * We must be careful about boundary cases so that we never
 625  * let the user have CTX 0 (nucleus) or we ever use a CTX
 626  * version of zero (and thus NO_CONTEXT would not be caught
 627  * by version mis-match tests in mmu_context.h).
 628  *
 629  * Always invoked with interrupts disabled.
 630  */
 631 void get_new_mmu_context(struct mm_struct *mm)
 632 {
 633         unsigned long ctx, new_ctx;
 634         unsigned long orig_pgsz_bits;
 635         unsigned long flags;
 636         int new_version;
 637
 638         spin_lock_irqsave(&ctx_alloc_lock, flags);
 639         orig_pgsz_bits = (mm->context.sparc64_ctx_val & CTX_PGSZ_MASK);
 640         ctx = (tlb_context_cache + 1) & CTX_NR_MASK;
 641         new_ctx = find_next_zero_bit(mmu_context_bmap, 1 << CTX_NR_BITS, ctx);
 642         new_version = 0;
 643         if (new_ctx >= (1 << CTX_NR_BITS)) {
 644                 new_ctx = find_next_zero_bit(mmu_context_bmap, ctx, 1);
 645                 if (new_ctx >= ctx) {
 646                         int i;
 647                         new_ctx = (tlb_context_cache & CTX_VERSION_MASK) +
 648                                 CTX_FIRST_VERSION;
 649                         if (new_ctx == 1)
 650                                 new_ctx = CTX_FIRST_VERSION;
 651
 652                         /* Don't call memset, for 16 entries that's just
 653                          * plain silly...
 654                          */
 655                         mmu_context_bmap[0] = 3;
 656                         mmu_context_bmap[1] = 0;
 657                         mmu_context_bmap[2] = 0;
 658                         mmu_context_bmap[3] = 0;
 659                         for (i = 4; i < CTX_BMAP_SLOTS; i += 4) {
 660                                 mmu_context_bmap[i + 0] = 0;
 661                                 mmu_context_bmap[i + 1] = 0;
 662                                 mmu_context_bmap[i + 2] = 0;
 663                                 mmu_context_bmap[i + 3] = 0;
 664                         }
 665                         new_version = 1;
 666                         goto out;
 667                 }
 668         }
 669         mmu_context_bmap[new_ctx>>6] |= (1UL << (new_ctx & 63));
 670         new_ctx |= (tlb_context_cache & CTX_VERSION_MASK);
 671 out:
 672         tlb_context_cache = new_ctx;
 673         mm->context.sparc64_ctx_val = new_ctx | orig_pgsz_bits;
 674         spin_unlock_irqrestore(&ctx_alloc_lock, flags);
 675
 676         if (unlikely(new_version))
 677                 smp_new_mmu_context_version();
 678 }
 679
 680 static int numa_enabled = 1;
 681 static int numa_debug;
 682
 683 static int __init early_numa(char *p)
 684 {
 685         if (!p)
 686                 return 0;
 687
 688         if (strstr(p, "off"))
 689                 numa_enabled = 0;
 690
 691         if (strstr(p, "debug"))
 692                 numa_debug = 1;
 693
 694         return 0;
 695 }
 696 early_param("numa", early_numa);
 697
 698 #define numadbg(f, a...) \
 699 do {    if (numa_debug) \
 700                 printk(KERN_INFO f, ## a); \
 701 } while (0)
 702
 703 static void __init find_ramdisk(unsigned long phys_base)
 704 {
 705 #ifdef CONFIG_BLK_DEV_INITRD
 706         if (sparc_ramdisk_image || sparc_ramdisk_image64) {
 707                 unsigned long ramdisk_image;
 708
 709                 /* Older versions of the bootloader only supported a
 710                  * 32-bit physical address for the ramdisk image
 711                  * location, stored at sparc_ramdisk_image.  Newer
 712                  * SILO versions set sparc_ramdisk_image to zero and
 713                  * provide a full 64-bit physical address at
 714                  * sparc_ramdisk_image64.
 715                  */
 716                 ramdisk_image = sparc_ramdisk_image;
 717                 if (!ramdisk_image)
 718                         ramdisk_image = sparc_ramdisk_image64;
 719
 720                 /* Another bootloader quirk.  The bootloader normalizes
 721                  * the physical address to KERNBASE, so we have to
 722                  * factor that back out and add in the lowest valid
 723                  * physical page address to get the true physical address.
 724                  */
 725                 ramdisk_image -= KERNBASE;
 726                 ramdisk_image += phys_base;
 727
 728                 numadbg("Found ramdisk at physical address 0x%lx, size %u\n",
 729                         ramdisk_image, sparc_ramdisk_size);
 730
 731                 initrd_start = ramdisk_image;
 732                 initrd_end = ramdisk_image + sparc_ramdisk_size;
 733
 734                 memblock_reserve(initrd_start, sparc_ramdisk_size);
 735
 736                 initrd_start += PAGE_OFFSET;
 737                 initrd_end += PAGE_OFFSET;
 738         }
 739 #endif
 740 }
 741
 742 struct node_mem_mask {
 743         unsigned long mask;
 744         unsigned long val;
 745         unsigned long bootmem_paddr;
 746 };
 747 static struct node_mem_mask node_masks[MAX_NUMNODES];
 748 static int num_node_masks;
 749
 750 int numa_cpu_lookup_table[NR_CPUS];
 751 cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
 752
 753 #ifdef CONFIG_NEED_MULTIPLE_NODES
 754
 755 struct mdesc_mblock {
 756         u64     base;
 757         u64     size;
 758         u64     offset; /* RA-to-PA */
 759 };
 760 static struct mdesc_mblock *mblocks;
 761 static int num_mblocks;
 762
 763 static unsigned long ra_to_pa(unsigned long addr)
 764 {
 765         int i;
 766
 767         for (i = 0; i < num_mblocks; i++) {
 768                 struct mdesc_mblock *m = &mblocks[i];
 769
 770                 if (addr >= m->base &&
 771                     addr < (m->base + m->size)) {
 772                         addr += m->offset;
 773                         break;
 774                 }
 775         }
 776         return addr;
 777 }
 778
 779 static int find_node(unsigned long addr)
 780 {
 781         int i;
 782
 783         addr = ra_to_pa(addr);
 784         for (i = 0; i < num_node_masks; i++) {
 785                 struct node_mem_mask *p = &node_masks[i];
 786
 787                 if ((addr & p->mask) == p->val)
 788                         return i;
 789         }
 790         return -1;
 791 }
 792
 793 u64 memblock_nid_range(u64 start, u64 end, int *nid)
 794 {
 795         *nid = find_node(start);
 796         start += PAGE_SIZE;
 797         while (start < end) {
 798                 int n = find_node(start);
 799
 800                 if (n != *nid)
 801                         break;
 802                 start += PAGE_SIZE;
 803         }
 804
 805         if (start > end)
 806                 start = end;
 807
 808         return start;
 809 }
 810 #else
 811 u64 memblock_nid_range(u64 start, u64 end, int *nid)
 812 {
 813         *nid = 0;
 814         return end;
 815 }
 816 #endif
 817
 818 /* This must be invoked after performing all of the necessary
 819  * add_active_range() calls for 'nid'.  We need to be able to get
 820  * correct data from get_pfn_range_for_nid().
 821  */
 822 static void __init allocate_node_data(int nid)
 823 {
 824         unsigned long paddr, num_pages, start_pfn, end_pfn;
 825         struct pglist_data *p;
 826
 827 #ifdef CONFIG_NEED_MULTIPLE_NODES
 828         paddr = memblock_alloc_try_nid(sizeof(struct pglist_data), SMP_CACHE_BYTES, nid);
 829         if (!paddr) {
 830                 prom_printf("Cannot allocate pglist_data for nid[%d]\n", nid);
 831                 prom_halt();
 832         }
 833         NODE_DATA(nid) = __va(paddr);
 834         memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
 835
 836         NODE_DATA(nid)->bdata = &bootmem_node_data[nid];
 837 #endif
 838
 839         p = NODE_DATA(nid);
 840
 841         get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
 842         p->node_start_pfn = start_pfn;
 843         p->node_spanned_pages = end_pfn - start_pfn;
 844
 845         if (p->node_spanned_pages) {
 846                 num_pages = bootmem_bootmap_pages(p->node_spanned_pages);
 847
 848                 paddr = memblock_alloc_try_nid(num_pages << PAGE_SHIFT, PAGE_SIZE, nid);
 849                 if (!paddr) {
 850                         prom_printf("Cannot allocate bootmap for nid[%d]\n",
 851                                   nid);
 852                         prom_halt();
 853                 }
 854                 node_masks[nid].bootmem_paddr = paddr;
 855         }
 856 }
 857
 858 static void init_node_masks_nonnuma(void)
 859 {
 860         int i;
 861
 862         numadbg("Initializing tables for non-numa.\n");
 863
 864         node_masks[0].mask = node_masks[0].val = 0;
 865         num_node_masks = 1;
 866
 867         for (i = 0; i < NR_CPUS; i++)
 868                 numa_cpu_lookup_table[i] = 0;
 869
 870         cpumask_setall(&numa_cpumask_lookup_table[0]);
 871 }
 872
 873 #ifdef CONFIG_NEED_MULTIPLE_NODES
 874 struct pglist_data *node_data[MAX_NUMNODES];
 875
 876 EXPORT_SYMBOL(numa_cpu_lookup_table);
 877 EXPORT_SYMBOL(numa_cpumask_lookup_table);
 878 EXPORT_SYMBOL(node_data);
 879
 880 struct mdesc_mlgroup {
 881         u64     node;
 882         u64     latency;
 883         u64     match;
 884         u64     mask;
 885 };
 886 static struct mdesc_mlgroup *mlgroups;
 887 static int num_mlgroups;
 888
 889 static int scan_pio_for_cfg_handle(struct mdesc_handle *md, u64 pio,
 890                                    u32 cfg_handle)
 891 {
 892         u64 arc;
 893
 894         mdesc_for_each_arc(arc, md, pio, MDESC_ARC_TYPE_FWD) {
 895                 u64 target = mdesc_arc_target(md, arc);
 896                 const u64 *val;
 897
 898                 val = mdesc_get_property(md, target,
 899                                          "cfg-handle", NULL);
 900                 if (val && *val == cfg_handle)
 901                         return 0;
 902         }
 903         return -ENODEV;
 904 }
 905
 906 static int scan_arcs_for_cfg_handle(struct mdesc_handle *md, u64 grp,
 907                                     u32 cfg_handle)
 908 {
 909         u64 arc, candidate, best_latency = ~(u64)0;
 910
 911         candidate = MDESC_NODE_NULL;
 912         mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
 913                 u64 target = mdesc_arc_target(md, arc);
 914                 const char *name = mdesc_node_name(md, target);
 915                 const u64 *val;
 916
 917                 if (strcmp(name, "pio-latency-group"))
 918                         continue;
 919
 920                 val = mdesc_get_property(md, target, "latency", NULL);
 921                 if (!val)
 922                         continue;
 923
 924                 if (*val < best_latency) {
 925                         candidate = target;
 926                         best_latency = *val;
 927                 }
 928         }
 929
 930         if (candidate == MDESC_NODE_NULL)
 931                 return -ENODEV;
 932
 933         return scan_pio_for_cfg_handle(md, candidate, cfg_handle);
 934 }
 935
 936 int of_node_to_nid(struct device_node *dp)
 937 {
 938         const struct linux_prom64_registers *regs;
 939         struct mdesc_handle *md;
 940         u32 cfg_handle;
 941         int count, nid;
 942         u64 grp;
 943
 944         /* This is the right thing to do on currently supported
 945          * SUN4U NUMA platforms as well, as the PCI controller does
 946          * not sit behind any particular memory controller.
 947          */
 948         if (!mlgroups)
 949                 return -1;
 950
 951         regs = of_get_property(dp, "reg", NULL);
 952         if (!regs)
 953                 return -1;
 954
 955         cfg_handle = (regs->phys_addr >> 32UL) & 0x0fffffff;
 956
 957         md = mdesc_grab();
 958
 959         count = 0;
 960         nid = -1;
 961         mdesc_for_each_node_by_name(md, grp, "group") {
 962                 if (!scan_arcs_for_cfg_handle(md, grp, cfg_handle)) {
 963                         nid = count;
 964                         break;
 965                 }
 966                 count++;
 967         }
 968
 969         mdesc_release(md);
 970
 971         return nid;
 972 }
 973
 974 static void __init add_node_ranges(void)
 975 {
 976         struct memblock_region *reg;
 977
 978         for_each_memblock(memory, reg) {
 979                 unsigned long size = reg->size;
 980                 unsigned long start, end;
 981
 982                 start = reg->base;
 983                 end = start + size;
 984                 while (start < end) {
 985                         unsigned long this_end;
 986                         int nid;
 987
 988                         this_end = memblock_nid_range(start, end, &nid);
 989
 990                         numadbg("Adding active range nid[%d] "
 991                                 "start[%lx] end[%lx]\n",
 992                                 nid, start, this_end);
 993
 994                         add_active_range(nid,
 995                                          start >> PAGE_SHIFT,
 996                                          this_end >> PAGE_SHIFT);
 997
 998                         start = this_end;
 999                 }
1000         }
1001 }
1002
1003 static int __init grab_mlgroups(struct mdesc_handle *md)
1004 {
1005         unsigned long paddr;
1006         int count = 0;
1007         u64 node;
1008
1009         mdesc_for_each_node_by_name(md, node, "memory-latency-group")
1010                 count++;
1011         if (!count)
1012                 return -ENOENT;
1013
1014         paddr = memblock_alloc(count * sizeof(struct mdesc_mlgroup),
1015                           SMP_CACHE_BYTES);
1016         if (!paddr)
1017                 return -ENOMEM;
1018
1019         mlgroups = __va(paddr);
1020         num_mlgroups = count;
1021
1022         count = 0;
1023         mdesc_for_each_node_by_name(md, node, "memory-latency-group") {
1024                 struct mdesc_mlgroup *m = &mlgroups[count++];
1025                 const u64 *val;
1026
1027                 m->node = node;
1028
1029                 val = mdesc_get_property(md, node, "latency", NULL);
1030                 m->latency = *val;
1031                 val = mdesc_get_property(md, node, "address-match", NULL);
1032                 m->match = *val;
1033                 val = mdesc_get_property(md, node, "address-mask", NULL);
1034                 m->mask = *val;
1035
1036                 numadbg("MLGROUP[%d]: node[%llx] latency[%llx] "
1037                         "match[%llx] mask[%llx]\n",
1038                         count - 1, m->node, m->latency, m->match, m->mask);
1039         }
1040
1041         return 0;
1042 }
1043
1044 static int __init grab_mblocks(struct mdesc_handle *md)
1045 {
1046         unsigned long paddr;
1047         int count = 0;
1048         u64 node;
1049
1050         mdesc_for_each_node_by_name(md, node, "mblock")
1051                 count++;
1052         if (!count)
1053                 return -ENOENT;
1054
1055         paddr = memblock_alloc(count * sizeof(struct mdesc_mblock),
1056                           SMP_CACHE_BYTES);
1057         if (!paddr)
1058                 return -ENOMEM;
1059
1060         mblocks = __va(paddr);
1061         num_mblocks = count;
1062
1063         count = 0;
1064         mdesc_for_each_node_by_name(md, node, "mblock") {
1065                 struct mdesc_mblock *m = &mblocks[count++];
1066                 const u64 *val;
1067
1068                 val = mdesc_get_property(md, node, "base", NULL);
1069                 m->base = *val;
1070                 val = mdesc_get_property(md, node, "size", NULL);
1071                 m->size = *val;
1072                 val = mdesc_get_property(md, node,
1073                                          "address-congruence-offset", NULL);
1074                 m->offset = *val;
1075
1076                 numadbg("MBLOCK[%d]: base[%llx] size[%llx] offset[%llx]\n",
1077                         count - 1, m->base, m->size, m->offset);
1078         }
1079
1080         return 0;
1081 }
1082
1083 static void __init numa_parse_mdesc_group_cpus(struct mdesc_handle *md,
1084                                                u64 grp, cpumask_t *mask)
1085 {
1086         u64 arc;
1087
1088         cpumask_clear(mask);
1089
1090         mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_BACK) {
1091                 u64 target = mdesc_arc_target(md, arc);
1092                 const char *name = mdesc_node_name(md, target);
1093                 const u64 *id;
1094
1095                 if (strcmp(name, "cpu"))
1096                         continue;
1097                 id = mdesc_get_property(md, target, "id", NULL);
1098                 if (*id < nr_cpu_ids)
1099                         cpumask_set_cpu(*id, mask);
1100         }
1101 }
1102
1103 static struct mdesc_mlgroup * __init find_mlgroup(u64 node)
1104 {
1105         int i;
1106
1107         for (i = 0; i < num_mlgroups; i++) {
1108                 struct mdesc_mlgroup *m = &mlgroups[i];
1109                 if (m->node == node)
1110                         return m;
1111         }
1112         return NULL;
1113 }
1114
1115 static int __init numa_attach_mlgroup(struct mdesc_handle *md, u64 grp,
1116                                       int index)
1117 {
1118         struct mdesc_mlgroup *candidate = NULL;
1119         u64 arc, best_latency = ~(u64)0;
1120         struct node_mem_mask *n;
1121
1122         mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1123                 u64 target = mdesc_arc_target(md, arc);
1124                 struct mdesc_mlgroup *m = find_mlgroup(target);
1125                 if (!m)
1126                         continue;
1127                 if (m->latency < best_latency) {
1128                         candidate = m;
1129                         best_latency = m->latency;
1130                 }
1131         }
1132         if (!candidate)
1133                 return -ENOENT;
1134
1135         if (num_node_masks != index) {
1136                 printk(KERN_ERR "Inconsistent NUMA state, "
1137                        "index[%d] != num_node_masks[%d]\n",
1138                        index, num_node_masks);
1139                 return -EINVAL;
1140         }
1141
1142         n = &node_masks[num_node_masks++];
1143
1144         n->mask = candidate->mask;
1145         n->val = candidate->match;
1146
1147         numadbg("NUMA NODE[%d]: mask[%lx] val[%lx] (latency[%llx])\n",
1148                 index, n->mask, n->val, candidate->latency);
1149
1150         return 0;
1151 }
1152
1153 static int __init numa_parse_mdesc_group(struct mdesc_handle *md, u64 grp,
1154                                          int index)
1155 {
1156         cpumask_t mask;
1157         int cpu;
1158
1159         numa_parse_mdesc_group_cpus(md, grp, &mask);
1160
1161         for_each_cpu(cpu, &mask)
1162                 numa_cpu_lookup_table[cpu] = index;
1163         cpumask_copy(&numa_cpumask_lookup_table[index], &mask);
1164
1165         if (numa_debug) {
1166                 printk(KERN_INFO "NUMA GROUP[%d]: cpus [ ", index);
1167                 for_each_cpu(cpu, &mask)
1168                         printk("%d ", cpu);
1169                 printk("]\n");
1170         }
1171
1172         return numa_attach_mlgroup(md, grp, index);
1173 }
1174
1175 static int __init numa_parse_mdesc(void)
1176 {
1177         struct mdesc_handle *md = mdesc_grab();
1178         int i, err, count;
1179         u64 node;
1180
1181         node = mdesc_node_by_name(md, MDESC_NODE_NULL, "latency-groups");
1182         if (node == MDESC_NODE_NULL) {
1183                 mdesc_release(md);
1184                 return -ENOENT;
1185         }
1186
1187         err = grab_mblocks(md);
1188         if (err < 0)
1189                 goto out;
1190
1191         err = grab_mlgroups(md);
1192         if (err < 0)
1193                 goto out;
1194
1195         count = 0;
1196         mdesc_for_each_node_by_name(md, node, "group") {
1197                 err = numa_parse_mdesc_group(md, node, count);
1198                 if (err < 0)
1199                         break;
1200                 count++;
1201         }
1202
1203         add_node_ranges();
1204
1205         for (i = 0; i < num_node_masks; i++) {
1206                 allocate_node_data(i);
1207                 node_set_online(i);
1208         }
1209
1210         err = 0;
1211 out:
1212         mdesc_release(md);
1213         return err;
1214 }
1215
1216 static int __init numa_parse_jbus(void)
1217 {
1218         unsigned long cpu, index;
1219
1220         /* NUMA node id is encoded in bits 36 and higher, and there is
1221          * a 1-to-1 mapping from CPU ID to NUMA node ID.
1222          */
1223         index = 0;
1224         for_each_present_cpu(cpu) {
1225                 numa_cpu_lookup_table[cpu] = index;
1226                 cpumask_copy(&numa_cpumask_lookup_table[index], cpumask_of(cpu));
1227                 node_masks[index].mask = ~((1UL << 36UL) - 1UL);
1228                 node_masks[index].val = cpu << 36UL;
1229
1230                 index++;
1231         }
1232         num_node_masks = index;
1233
1234         add_node_ranges();
1235
1236         for (index = 0; index < num_node_masks; index++) {
1237                 allocate_node_data(index);
1238                 node_set_online(index);
1239         }
1240
1241         return 0;
1242 }
1243
1244 static int __init numa_parse_sun4u(void)
1245 {
1246         if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1247                 unsigned long ver;
1248
1249                 __asm__ ("rdpr %%ver, %0" : "=r" (ver));
1250                 if ((ver >> 32UL) == __JALAPENO_ID ||
1251                     (ver >> 32UL) == __SERRANO_ID)
1252                         return numa_parse_jbus();
1253         }
1254         return -1;
1255 }
1256
1257 static int __init bootmem_init_numa(void)
1258 {
1259         int err = -1;
1260
1261         numadbg("bootmem_init_numa()\n");
1262
1263         if (numa_enabled) {
1264                 if (tlb_type == hypervisor)
1265                         err = numa_parse_mdesc();
1266                 else
1267                         err = numa_parse_sun4u();
1268         }
1269         return err;
1270 }
1271
1272 #else
1273
1274 static int bootmem_init_numa(void)
1275 {
1276         return -1;
1277 }
1278
1279 #endif
1280
1281 static void __init bootmem_init_nonnuma(void)
1282 {
1283         unsigned long top_of_ram = memblock_end_of_DRAM();
1284         unsigned long total_ram = memblock_phys_mem_size();
1285         struct memblock_region *reg;
1286
1287         numadbg("bootmem_init_nonnuma()\n");
1288
1289         printk(KERN_INFO "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
1290                top_of_ram, total_ram);
1291         printk(KERN_INFO "Memory hole size: %ldMB\n",
1292                (top_of_ram - total_ram) >> 20);
1293
1294         init_node_masks_nonnuma();
1295
1296         for_each_memblock(memory, reg) {
1297                 unsigned long start_pfn, end_pfn;
1298
1299                 if (!reg->size)
1300                         continue;
1301
1302                 start_pfn = memblock_region_memory_base_pfn(reg);
1303                 end_pfn = memblock_region_memory_end_pfn(reg);
1304                 add_active_range(0, start_pfn, end_pfn);
1305         }
1306
1307         allocate_node_data(0);
1308
1309         node_set_online(0);
1310 }
1311
1312 static void __init reserve_range_in_node(int nid, unsigned long start,
1313                                          unsigned long end)
1314 {
1315         numadbg("    reserve_range_in_node(nid[%d],start[%lx],end[%lx]\n",
1316                 nid, start, end);
1317         while (start < end) {
1318                 unsigned long this_end;
1319                 int n;
1320
1321                 this_end = memblock_nid_range(start, end, &n);
1322                 if (n == nid) {
1323                         numadbg("      MATCH reserving range [%lx:%lx]\n",
1324                                 start, this_end);
1325                         reserve_bootmem_node(NODE_DATA(nid), start,
1326                                              (this_end - start), BOOTMEM_DEFAULT);
1327                 } else
1328                         numadbg("      NO MATCH, advancing start to %lx\n",
1329                                 this_end);
1330
1331                 start = this_end;
1332         }
1333 }
1334
1335 static void __init trim_reserved_in_node(int nid)
1336 {
1337         struct memblock_region *reg;
1338
1339         numadbg("  trim_reserved_in_node(%d)\n", nid);
1340
1341         for_each_memblock(reserved, reg)
1342                 reserve_range_in_node(nid, reg->base, reg->base + reg->size);
1343 }
1344
1345 static void __init bootmem_init_one_node(int nid)
1346 {
1347         struct pglist_data *p;
1348
1349         numadbg("bootmem_init_one_node(%d)\n", nid);
1350
1351         p = NODE_DATA(nid);
1352
1353         if (p->node_spanned_pages) {
1354                 unsigned long paddr = node_masks[nid].bootmem_paddr;
1355                 unsigned long end_pfn;
1356
1357                 end_pfn = p->node_start_pfn + p->node_spanned_pages;
1358
1359                 numadbg("  init_bootmem_node(%d, %lx, %lx, %lx)\n",
1360                         nid, paddr >> PAGE_SHIFT, p->node_start_pfn, end_pfn);
1361
1362                 init_bootmem_node(p, paddr >> PAGE_SHIFT,
1363                                   p->node_start_pfn, end_pfn);
1364
1365                 numadbg("  free_bootmem_with_active_regions(%d, %lx)\n",
1366                         nid, end_pfn);
1367                 free_bootmem_with_active_regions(nid, end_pfn);
1368
1369                 trim_reserved_in_node(nid);
1370
1371                 numadbg("  sparse_memory_present_with_active_regions(%d)\n",
1372                         nid);
1373                 sparse_memory_present_with_active_regions(nid);
1374         }
1375 }
1376
1377 static unsigned long __init bootmem_init(unsigned long phys_base)
1378 {
1379         unsigned long end_pfn;
1380         int nid;
1381
1382         end_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1383         max_pfn = max_low_pfn = end_pfn;
1384         min_low_pfn = (phys_base >> PAGE_SHIFT);
1385
1386         if (bootmem_init_numa() < 0)
1387                 bootmem_init_nonnuma();
1388
1389         /* XXX cpu notifier XXX */
1390
1391         for_each_online_node(nid)
1392                 bootmem_init_one_node(nid);
1393
1394         sparse_init();
1395
1396         return end_pfn;
1397 }
1398
1399 static struct linux_prom64_registers pall[MAX_BANKS] __initdata;
1400 static int pall_ents __initdata;
1401
1402 #ifdef CONFIG_DEBUG_PAGEALLOC
1403 static unsigned long __ref kernel_map_range(unsigned long pstart,
1404                                             unsigned long pend, pgprot_t prot)
1405 {
1406         unsigned long vstart = PAGE_OFFSET + pstart;
1407         unsigned long vend = PAGE_OFFSET + pend;
1408         unsigned long alloc_bytes = 0UL;
1409
1410         if ((vstart & ~PAGE_MASK) || (vend & ~PAGE_MASK)) {
1411                 prom_printf("kernel_map: Unaligned physmem[%lx:%lx]\n",
1412                             vstart, vend);
1413                 prom_halt();
1414         }
1415
1416         while (vstart < vend) {
1417                 unsigned long this_end, paddr = __pa(vstart);
1418                 pgd_t *pgd = pgd_offset_k(vstart);
1419                 pud_t *pud;
1420                 pmd_t *pmd;
1421                 pte_t *pte;
1422
1423                 pud = pud_offset(pgd, vstart);
1424                 if (pud_none(*pud)) {
1425                         pmd_t *new;
1426
1427                         new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1428                         alloc_bytes += PAGE_SIZE;
1429                         pud_populate(&init_mm, pud, new);
1430                 }
1431
1432                 pmd = pmd_offset(pud, vstart);
1433                 if (!pmd_present(*pmd)) {
1434                         pte_t *new;
1435
1436                         new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1437                         alloc_bytes += PAGE_SIZE;
1438                         pmd_populate_kernel(&init_mm, pmd, new);
1439                 }
1440
1441                 pte = pte_offset_kernel(pmd, vstart);
1442                 this_end = (vstart + PMD_SIZE) & PMD_MASK;
1443                 if (this_end > vend)
1444                         this_end = vend;
1445
1446                 while (vstart < this_end) {
1447                         pte_val(*pte) = (paddr | pgprot_val(prot));
1448
1449                         vstart += PAGE_SIZE;
1450                         paddr += PAGE_SIZE;
1451                         pte++;
1452                 }
1453         }
1454
1455         return alloc_bytes;
1456 }
1457
1458 extern unsigned int kvmap_linear_patch[1];
1459 #endif /* CONFIG_DEBUG_PAGEALLOC */
1460
1461 static void __init mark_kpte_bitmap(unsigned long start, unsigned long end)
1462 {
1463         const unsigned long shift_256MB = 28;
1464         const unsigned long mask_256MB = ((1UL << shift_256MB) - 1UL);
1465         const unsigned long size_256MB = (1UL << shift_256MB);
1466
1467         while (start < end) {
1468                 long remains;
1469
1470                 remains = end - start;
1471                 if (remains < size_256MB)
1472                         break;
1473
1474                 if (start & mask_256MB) {
1475                         start = (start + size_256MB) & ~mask_256MB;
1476                         continue;
1477                 }
1478
1479                 while (remains >= size_256MB) {
1480                         unsigned long index = start >> shift_256MB;
1481
1482                         __set_bit(index, kpte_linear_bitmap);
1483
1484                         start += size_256MB;
1485                         remains -= size_256MB;
1486                 }
1487         }
1488 }
1489
1490 static void __init init_kpte_bitmap(void)
1491 {
1492         unsigned long i;
1493
1494         for (i = 0; i < pall_ents; i++) {
1495                 unsigned long phys_start, phys_end;
1496
1497                 phys_start = pall[i].phys_addr;
1498                 phys_end = phys_start + pall[i].reg_size;
1499
1500                 mark_kpte_bitmap(phys_start, phys_end);
1501         }
1502 }
1503
1504 static void __init kernel_physical_mapping_init(void)
1505 {
1506 #ifdef CONFIG_DEBUG_PAGEALLOC
1507         unsigned long i, mem_alloced = 0UL;
1508
1509         for (i = 0; i < pall_ents; i++) {
1510                 unsigned long phys_start, phys_end;
1511
1512                 phys_start = pall[i].phys_addr;
1513                 phys_end = phys_start + pall[i].reg_size;
1514
1515                 mem_alloced += kernel_map_range(phys_start, phys_end,
1516                                                 PAGE_KERNEL);
1517         }
1518
1519         printk("Allocated %ld bytes for kernel page tables.\n",
1520                mem_alloced);
1521
1522         kvmap_linear_patch[0] = 0x01000000; /* nop */
1523         flushi(&kvmap_linear_patch[0]);
1524
1525         __flush_tlb_all();
1526 #endif
1527 }
1528
1529 #ifdef CONFIG_DEBUG_PAGEALLOC
1530 void kernel_map_pages(struct page *page, int numpages, int enable)
1531 {
1532         unsigned long phys_start = page_to_pfn(page) << PAGE_SHIFT;
1533         unsigned long phys_end = phys_start + (numpages * PAGE_SIZE);
1534
1535         kernel_map_range(phys_start, phys_end,
1536                          (enable ? PAGE_KERNEL : __pgprot(0)));
1537
1538         flush_tsb_kernel_range(PAGE_OFFSET + phys_start,
1539                                PAGE_OFFSET + phys_end);
1540
1541         /* we should perform an IPI and flush all tlbs,
1542          * but that can deadlock->flush only current cpu.
1543          */
1544         __flush_tlb_kernel_range(PAGE_OFFSET + phys_start,
1545                                  PAGE_OFFSET + phys_end);
1546 }
1547 #endif
1548
1549 unsigned long __init find_ecache_flush_span(unsigned long size)
1550 {
1551         int i;
1552
1553         for (i = 0; i < pavail_ents; i++) {
1554                 if (pavail[i].reg_size >= size)
1555                         return pavail[i].phys_addr;
1556         }
1557
1558         return ~0UL;
1559 }
1560
1561 static void __init tsb_phys_patch(void)
1562 {
1563         struct tsb_ldquad_phys_patch_entry *pquad;
1564         struct tsb_phys_patch_entry *p;
1565
1566         pquad = &__tsb_ldquad_phys_patch;
1567         while (pquad < &__tsb_ldquad_phys_patch_end) {
1568                 unsigned long addr = pquad->addr;
1569
1570                 if (tlb_type == hypervisor)
1571                         *(unsigned int *) addr = pquad->sun4v_insn;
1572                 else
1573                         *(unsigned int *) addr = pquad->sun4u_insn;
1574                 wmb();
1575                 __asm__ __volatile__("flush     %0"
1576                                      : /* no outputs */
1577                                      : "r" (addr));
1578
1579                 pquad++;
1580         }
1581
1582         p = &__tsb_phys_patch;
1583         while (p < &__tsb_phys_patch_end) {
1584                 unsigned long addr = p->addr;
1585
1586                 *(unsigned int *) addr = p->insn;
1587                 wmb();
1588                 __asm__ __volatile__("flush     %0"
1589                                      : /* no outputs */
1590                                      : "r" (addr));
1591
1592                 p++;
1593         }
1594 }
1595
1596 /* Don't mark as init, we give this to the Hypervisor.  */
1597 #ifndef CONFIG_DEBUG_PAGEALLOC
1598 #define NUM_KTSB_DESCR  2
1599 #else
1600 #define NUM_KTSB_DESCR  1
1601 #endif
1602 static struct hv_tsb_descr ktsb_descr[NUM_KTSB_DESCR];
1603 extern struct tsb swapper_tsb[KERNEL_TSB_NENTRIES];
1604
1605 static void patch_one_ktsb_phys(unsigned int *start, unsigned int *end, unsigned long pa)
1606 {
1607         pa >>= KTSB_PHYS_SHIFT;
1608
1609         while (start < end) {
1610                 unsigned int *ia = (unsigned int *)(unsigned long)*start;
1611
1612                 ia[0] = (ia[0] & ~0x3fffff) | (pa >> 10);
1613                 __asm__ __volatile__("flush     %0" : : "r" (ia));
1614
1615                 ia[1] = (ia[1] & ~0x3ff) | (pa & 0x3ff);
1616                 __asm__ __volatile__("flush     %0" : : "r" (ia + 1));
1617
1618                 start++;
1619         }
1620 }
1621
1622 static void ktsb_phys_patch(void)
1623 {
1624         extern unsigned int __swapper_tsb_phys_patch;
1625         extern unsigned int __swapper_tsb_phys_patch_end;
1626         unsigned long ktsb_pa;
1627
1628         ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
1629         patch_one_ktsb_phys(&__swapper_tsb_phys_patch,
1630                             &__swapper_tsb_phys_patch_end, ktsb_pa);
1631 #ifndef CONFIG_DEBUG_PAGEALLOC
1632         {
1633         extern unsigned int __swapper_4m_tsb_phys_patch;
1634         extern unsigned int __swapper_4m_tsb_phys_patch_end;
1635         ktsb_pa = (kern_base +
1636                    ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
1637         patch_one_ktsb_phys(&__swapper_4m_tsb_phys_patch,
1638                             &__swapper_4m_tsb_phys_patch_end, ktsb_pa);
1639         }
1640 #endif
1641 }
1642
1643 static void __init sun4v_ktsb_init(void)
1644 {
1645         unsigned long ktsb_pa;
1646
1647         /* First KTSB for PAGE_SIZE mappings.  */
1648         ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
1649
1650         switch (PAGE_SIZE) {
1651         case 8 * 1024:
1652         default:
1653                 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_8K;
1654                 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_8K;
1655                 break;
1656
1657         case 64 * 1024:
1658                 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_64K;
1659                 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_64K;
1660                 break;
1661
1662         case 512 * 1024:
1663                 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_512K;
1664                 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_512K;
1665                 break;
1666
1667         case 4 * 1024 * 1024:
1668                 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_4MB;
1669                 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_4MB;
1670                 break;
1671         }
1672
1673         ktsb_descr[0].assoc = 1;
1674         ktsb_descr[0].num_ttes = KERNEL_TSB_NENTRIES;
1675         ktsb_descr[0].ctx_idx = 0;
1676         ktsb_descr[0].tsb_base = ktsb_pa;
1677         ktsb_descr[0].resv = 0;
1678
1679 #ifndef CONFIG_DEBUG_PAGEALLOC
1680         /* Second KTSB for 4MB/256MB mappings.  */
1681         ktsb_pa = (kern_base +
1682                    ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
1683
1684         ktsb_descr[1].pgsz_idx = HV_PGSZ_IDX_4MB;
1685         ktsb_descr[1].pgsz_mask = (HV_PGSZ_MASK_4MB |
1686                                    HV_PGSZ_MASK_256MB);
1687         ktsb_descr[1].assoc = 1;
1688         ktsb_descr[1].num_ttes = KERNEL_TSB4M_NENTRIES;
1689         ktsb_descr[1].ctx_idx = 0;
1690         ktsb_descr[1].tsb_base = ktsb_pa;
1691         ktsb_descr[1].resv = 0;
1692 #endif
1693 }
1694
1695 void __cpuinit sun4v_ktsb_register(void)
1696 {
1697         unsigned long pa, ret;
1698
1699         pa = kern_base + ((unsigned long)&ktsb_descr[0] - KERNBASE);
1700
1701         ret = sun4v_mmu_tsb_ctx0(NUM_KTSB_DESCR, pa);
1702         if (ret != 0) {
1703                 prom_printf("hypervisor_mmu_tsb_ctx0[%lx]: "
1704                             "errors with %lx\n", pa, ret);
1705                 prom_halt();
1706         }
1707 }
1708
1709 /* paging_init() sets up the page tables */
1710
1711 static unsigned long last_valid_pfn;
1712 pgd_t swapper_pg_dir[2048];
1713
1714 static void sun4u_pgprot_init(void);
1715 static void sun4v_pgprot_init(void);
1716
1717 void __init paging_init(void)
1718 {
1719         unsigned long end_pfn, shift, phys_base;
1720         unsigned long real_end, i;
1721
1722         /* These build time checkes make sure that the dcache_dirty_cpu()
1723          * page->flags usage will work.
1724          *
1725          * When a page gets marked as dcache-dirty, we store the
1726          * cpu number starting at bit 32 in the page->flags.  Also,
1727          * functions like clear_dcache_dirty_cpu use the cpu mask
1728          * in 13-bit signed-immediate instruction fields.
1729          */
1730
1731         /*
1732          * Page flags must not reach into upper 32 bits that are used
1733          * for the cpu number
1734          */
1735         BUILD_BUG_ON(NR_PAGEFLAGS > 32);
1736
1737         /*
1738          * The bit fields placed in the high range must not reach below
1739          * the 32 bit boundary. Otherwise we cannot place the cpu field
1740          * at the 32 bit boundary.
1741          */
1742         BUILD_BUG_ON(SECTIONS_WIDTH + NODES_WIDTH + ZONES_WIDTH +
1743                 ilog2(roundup_pow_of_two(NR_CPUS)) > 32);
1744
1745         BUILD_BUG_ON(NR_CPUS > 4096);
1746
1747         kern_base = (prom_boot_mapping_phys_low >> 22UL) << 22UL;
1748         kern_size = (unsigned long)&_end - (unsigned long)KERNBASE;
1749
1750         /* Invalidate both kernel TSBs.  */
1751         memset(swapper_tsb, 0x40, sizeof(swapper_tsb));
1752 #ifndef CONFIG_DEBUG_PAGEALLOC
1753         memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
1754 #endif
1755
1756         if (tlb_type == hypervisor)
1757                 sun4v_pgprot_init();
1758         else
1759                 sun4u_pgprot_init();
1760
1761         if (tlb_type == cheetah_plus ||
1762             tlb_type == hypervisor) {
1763                 tsb_phys_patch();
1764                 ktsb_phys_patch();
1765         }
1766
1767         if (tlb_type == hypervisor) {
1768                 sun4v_patch_tlb_handlers();
1769                 sun4v_ktsb_init();
1770         }
1771
1772         memblock_init();
1773
1774         /* Find available physical memory...
1775          *
1776          * Read it twice in order to work around a bug in openfirmware.
1777          * The call to grab this table itself can cause openfirmware to
1778          * allocate memory, which in turn can take away some space from
1779          * the list of available memory.  Reading it twice makes sure
1780          * we really do get the final value.
1781          */
1782         read_obp_translations();
1783         read_obp_memory("reg", &pall[0], &pall_ents);
1784         read_obp_memory("available", &pavail[0], &pavail_ents);
1785         read_obp_memory("available", &pavail[0], &pavail_ents);
1786
1787         phys_base = 0xffffffffffffffffUL;
1788         for (i = 0; i < pavail_ents; i++) {
1789                 phys_base = min(phys_base, pavail[i].phys_addr);
1790                 memblock_add(pavail[i].phys_addr, pavail[i].reg_size);
1791         }
1792
1793         memblock_reserve(kern_base, kern_size);
1794
1795         find_ramdisk(phys_base);
1796
1797         memblock_enforce_memory_limit(cmdline_memory_size);
1798
1799         memblock_analyze();
1800         memblock_dump_all();
1801
1802         set_bit(0, mmu_context_bmap);
1803
1804         shift = kern_base + PAGE_OFFSET - ((unsigned long)KERNBASE);
1805
1806         real_end = (unsigned long)_end;
1807         num_kernel_image_mappings = DIV_ROUND_UP(real_end - KERNBASE, 1 << 22);
1808         printk("Kernel: Using %d locked TLB entries for main kernel image.\n",
1809                num_kernel_image_mappings);
1810
1811         /* Set kernel pgd to upper alias so physical page computations
1812          * work.
1813          */
1814         init_mm.pgd += ((shift) / (sizeof(pgd_t)));
1815
1816         memset(swapper_low_pmd_dir, 0, sizeof(swapper_low_pmd_dir));
1817
1818         /* Now can init the kernel/bad page tables. */
1819         pud_set(pud_offset(&swapper_pg_dir[0], 0),
1820                 swapper_low_pmd_dir + (shift / sizeof(pgd_t)));
1821
1822         inherit_prom_mappings();
1823
1824         init_kpte_bitmap();
1825
1826         /* Ok, we can use our TLB miss and window trap handlers safely.  */
1827         setup_tba();
1828
1829         __flush_tlb_all();
1830
1831         if (tlb_type == hypervisor)
1832                 sun4v_ktsb_register();
1833
1834         prom_build_devicetree();
1835         of_populate_present_mask();
1836 #ifndef CONFIG_SMP
1837         of_fill_in_cpu_data();
1838 #endif
1839
1840         if (tlb_type == hypervisor) {
1841                 sun4v_mdesc_init();
1842                 mdesc_populate_present_mask(cpu_all_mask);
1843 #ifndef CONFIG_SMP
1844                 mdesc_fill_in_cpu_data(cpu_all_mask);
1845 #endif
1846         }
1847
1848         /* Once the OF device tree and MDESC have been setup, we know
1849          * the list of possible cpus.  Therefore we can allocate the
1850          * IRQ stacks.
1851          */
1852         for_each_possible_cpu(i) {
1853                 /* XXX Use node local allocations... XXX */
1854                 softirq_stack[i] = __va(memblock_alloc(THREAD_SIZE, THREAD_SIZE));
1855                 hardirq_stack[i] = __va(memblock_alloc(THREAD_SIZE, THREAD_SIZE));
1856         }
1857
1858         /* Setup bootmem... */
1859         last_valid_pfn = end_pfn = bootmem_init(phys_base);
1860
1861 #ifndef CONFIG_NEED_MULTIPLE_NODES
1862         max_mapnr = last_valid_pfn;
1863 #endif
1864         kernel_physical_mapping_init();
1865
1866         {
1867                 unsigned long max_zone_pfns[MAX_NR_ZONES];
1868
1869                 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
1870
1871                 max_zone_pfns[ZONE_NORMAL] = end_pfn;
1872
1873                 free_area_init_nodes(max_zone_pfns);
1874         }
1875
1876         printk("Booting Linux...\n");
1877 }
1878
1879 int __devinit page_in_phys_avail(unsigned long paddr)
1880 {
1881         int i;
1882
1883         paddr &= PAGE_MASK;
1884
1885         for (i = 0; i < pavail_ents; i++) {
1886                 unsigned long start, end;
1887
1888                 start = pavail[i].phys_addr;
1889                 end = start + pavail[i].reg_size;
1890
1891                 if (paddr >= start && paddr < end)
1892                         return 1;
1893         }
1894         if (paddr >= kern_base && paddr < (kern_base + kern_size))
1895                 return 1;
1896 #ifdef CONFIG_BLK_DEV_INITRD
1897         if (paddr >= __pa(initrd_start) &&
1898             paddr < __pa(PAGE_ALIGN(initrd_end)))
1899                 return 1;
1900 #endif
1901
1902         return 0;
1903 }
1904
1905 static struct linux_prom64_registers pavail_rescan[MAX_BANKS] __initdata;
1906 static int pavail_rescan_ents __initdata;
1907
1908 /* Certain OBP calls, such as fetching "available" properties, can
1909  * claim physical memory.  So, along with initializing the valid
1910  * address bitmap, what we do here is refetch the physical available
1911  * memory list again, and make sure it provides at least as much
1912  * memory as 'pavail' does.
1913  */
1914 static void __init setup_valid_addr_bitmap_from_pavail(unsigned long *bitmap)
1915 {
1916         int i;
1917
1918         read_obp_memory("available", &pavail_rescan[0], &pavail_rescan_ents);
1919
1920         for (i = 0; i < pavail_ents; i++) {
1921                 unsigned long old_start, old_end;
1922
1923                 old_start = pavail[i].phys_addr;
1924                 old_end = old_start + pavail[i].reg_size;
1925                 while (old_start < old_end) {
1926                         int n;
1927
1928                         for (n = 0; n < pavail_rescan_ents; n++) {
1929                                 unsigned long new_start, new_end;
1930
1931                                 new_start = pavail_rescan[n].phys_addr;
1932                                 new_end = new_start +
1933                                         pavail_rescan[n].reg_size;
1934
1935                                 if (new_start <= old_start &&
1936                                     new_end >= (old_start + PAGE_SIZE)) {
1937                                         set_bit(old_start >> 22, bitmap);
1938                                         goto do_next_page;
1939                                 }
1940                         }
1941
1942                         prom_printf("mem_init: Lost memory in pavail\n");
1943                         prom_printf("mem_init: OLD start[%lx] size[%lx]\n",
1944                                     pavail[i].phys_addr,
1945                                     pavail[i].reg_size);
1946                         prom_printf("mem_init: NEW start[%lx] size[%lx]\n",
1947                                     pavail_rescan[i].phys_addr,
1948                                     pavail_rescan[i].reg_size);
1949                         prom_printf("mem_init: Cannot continue, aborting.\n");
1950                         prom_halt();
1951
1952                 do_next_page:
1953                         old_start += PAGE_SIZE;
1954                 }
1955         }
1956 }
1957
1958 static void __init patch_tlb_miss_handler_bitmap(void)
1959 {
1960         extern unsigned int valid_addr_bitmap_insn[];
1961         extern unsigned int valid_addr_bitmap_patch[];
1962
1963         valid_addr_bitmap_insn[1] = valid_addr_bitmap_patch[1];
1964         mb();
1965         valid_addr_bitmap_insn[0] = valid_addr_bitmap_patch[0];
1966         flushi(&valid_addr_bitmap_insn[0]);
1967 }
1968
1969 void __init mem_init(void)
1970 {
1971         unsigned long codepages, datapages, initpages;
1972         unsigned long addr, last;
1973
1974         addr = PAGE_OFFSET + kern_base;
1975         last = PAGE_ALIGN(kern_size) + addr;
1976         while (addr < last) {
1977                 set_bit(__pa(addr) >> 22, sparc64_valid_addr_bitmap);
1978                 addr += PAGE_SIZE;
1979         }
1980
1981         setup_valid_addr_bitmap_from_pavail(sparc64_valid_addr_bitmap);
1982         patch_tlb_miss_handler_bitmap();
1983
1984         high_memory = __va(last_valid_pfn << PAGE_SHIFT);
1985
1986 #ifdef CONFIG_NEED_MULTIPLE_NODES
1987         {
1988                 int i;
1989                 for_each_online_node(i) {
1990                         if (NODE_DATA(i)->node_spanned_pages != 0) {
1991                                 totalram_pages +=
1992                                         free_all_bootmem_node(NODE_DATA(i));
1993                         }
1994                 }
1995         }
1996 #else
1997         totalram_pages = free_all_bootmem();
1998 #endif
1999
2000         /* We subtract one to account for the mem_map_zero page
2001          * allocated below.
2002          */
2003         totalram_pages -= 1;
2004         num_physpages = totalram_pages;
2005
2006         /*
2007          * Set up the zero page, mark it reserved, so that page count
2008          * is not manipulated when freeing the page from user ptes.
2009          */
2010         mem_map_zero = alloc_pages(GFP_KERNEL|__GFP_ZERO, 0);
2011         if (mem_map_zero == NULL) {
2012                 prom_printf("paging_init: Cannot alloc zero page.\n");
2013                 prom_halt();
2014         }
2015         SetPageReserved(mem_map_zero);
2016
2017         codepages = (((unsigned long) _etext) - ((unsigned long) _start));
2018         codepages = PAGE_ALIGN(codepages) >> PAGE_SHIFT;
2019         datapages = (((unsigned long) _edata) - ((unsigned long) _etext));
2020         datapages = PAGE_ALIGN(datapages) >> PAGE_SHIFT;
2021         initpages = (((unsigned long) __init_end) - ((unsigned long) __init_begin));
2022         initpages = PAGE_ALIGN(initpages) >> PAGE_SHIFT;
2023
2024         printk("Memory: %luk available (%ldk kernel code, %ldk data, %ldk init) [%016lx,%016lx]\n",
2025                nr_free_pages() << (PAGE_SHIFT-10),
2026                codepages << (PAGE_SHIFT-10),
2027                datapages << (PAGE_SHIFT-10),
2028                initpages << (PAGE_SHIFT-10),
2029                PAGE_OFFSET, (last_valid_pfn << PAGE_SHIFT));
2030
2031         if (tlb_type == cheetah || tlb_type == cheetah_plus)
2032                 cheetah_ecache_flush_init();
2033 }
2034
2035 void free_initmem(void)
2036 {
2037         unsigned long addr, initend;
2038         int do_free = 1;
2039
2040         /* If the physical memory maps were trimmed by kernel command
2041          * line options, don't even try freeing this initmem stuff up.
2042          * The kernel image could have been in the trimmed out region
2043          * and if so the freeing below will free invalid page structs.
2044          */
2045         if (cmdline_memory_size)
2046                 do_free = 0;
2047
2048         /*
2049          * The init section is aligned to 8k in vmlinux.lds. Page align for >8k pagesizes.
2050          */
2051         addr = PAGE_ALIGN((unsigned long)(__init_begin));
2052         initend = (unsigned long)(__init_end) & PAGE_MASK;
2053         for (; addr < initend; addr += PAGE_SIZE) {
2054                 unsigned long page;
2055                 struct page *p;
2056
2057                 page = (addr +
2058                         ((unsigned long) __va(kern_base)) -
2059                         ((unsigned long) KERNBASE));
2060                 memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
2061
2062                 if (do_free) {
2063                         p = virt_to_page(page);
2064
2065                         ClearPageReserved(p);
2066                         init_page_count(p);
2067                         __free_page(p);
2068                         num_physpages++;
2069                         totalram_pages++;
2070                 }
2071         }
2072 }
2073
2074 #ifdef CONFIG_BLK_DEV_INITRD
2075 void free_initrd_mem(unsigned long start, unsigned long end)
2076 {
2077         if (start < end)
2078                 printk ("Freeing initrd memory: %ldk freed\n", (end - start) >> 10);
2079         for (; start < end; start += PAGE_SIZE) {
2080                 struct page *p = virt_to_page(start);
2081
2082                 ClearPageReserved(p);
2083                 init_page_count(p);
2084                 __free_page(p);
2085                 num_physpages++;
2086                 totalram_pages++;
2087         }
2088 }
2089 #endif
2090
2091 #define _PAGE_CACHE_4U  (_PAGE_CP_4U | _PAGE_CV_4U)
2092 #define _PAGE_CACHE_4V  (_PAGE_CP_4V | _PAGE_CV_4V)
2093 #define __DIRTY_BITS_4U  (_PAGE_MODIFIED_4U | _PAGE_WRITE_4U | _PAGE_W_4U)
2094 #define __DIRTY_BITS_4V  (_PAGE_MODIFIED_4V | _PAGE_WRITE_4V | _PAGE_W_4V)
2095 #define __ACCESS_BITS_4U (_PAGE_ACCESSED_4U | _PAGE_READ_4U | _PAGE_R)
2096 #define __ACCESS_BITS_4V (_PAGE_ACCESSED_4V | _PAGE_READ_4V | _PAGE_R)
2097
2098 pgprot_t PAGE_KERNEL __read_mostly;
2099 EXPORT_SYMBOL(PAGE_KERNEL);
2100
2101 pgprot_t PAGE_KERNEL_LOCKED __read_mostly;
2102 pgprot_t PAGE_COPY __read_mostly;
2103
2104 pgprot_t PAGE_SHARED __read_mostly;
2105 EXPORT_SYMBOL(PAGE_SHARED);
2106
2107 unsigned long pg_iobits __read_mostly;
2108
2109 unsigned long _PAGE_IE __read_mostly;
2110 EXPORT_SYMBOL(_PAGE_IE);
2111
2112 unsigned long _PAGE_E __read_mostly;
2113 EXPORT_SYMBOL(_PAGE_E);
2114
2115 unsigned long _PAGE_CACHE __read_mostly;
2116 EXPORT_SYMBOL(_PAGE_CACHE);
2117
2118 #ifdef CONFIG_SPARSEMEM_VMEMMAP
2119 unsigned long vmemmap_table[VMEMMAP_SIZE];
2120
2121 int __meminit vmemmap_populate(struct page *start, unsigned long nr, int node)
2122 {
2123         unsigned long vstart = (unsigned long) start;
2124         unsigned long vend = (unsigned long) (start + nr);
2125         unsigned long phys_start = (vstart - VMEMMAP_BASE);
2126         unsigned long phys_end = (vend - VMEMMAP_BASE);
2127         unsigned long addr = phys_start & VMEMMAP_CHUNK_MASK;
2128         unsigned long end = VMEMMAP_ALIGN(phys_end);
2129         unsigned long pte_base;
2130
2131         pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2132                     _PAGE_CP_4U | _PAGE_CV_4U |
2133                     _PAGE_P_4U | _PAGE_W_4U);
2134         if (tlb_type == hypervisor)
2135                 pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2136                             _PAGE_CP_4V | _PAGE_CV_4V |
2137                             _PAGE_P_4V | _PAGE_W_4V);
2138
2139         for (; addr < end; addr += VMEMMAP_CHUNK) {
2140                 unsigned long *vmem_pp =
2141                         vmemmap_table + (addr >> VMEMMAP_CHUNK_SHIFT);
2142                 void *block;
2143
2144                 if (!(*vmem_pp & _PAGE_VALID)) {
2145                         block = vmemmap_alloc_block(1UL << 22, node);
2146                         if (!block)
2147                                 return -ENOMEM;
2148
2149                         *vmem_pp = pte_base | __pa(block);
2150
2151                         printk(KERN_INFO "[%p-%p] page_structs=%lu "
2152                                "node=%d entry=%lu/%lu\n", start, block, nr,
2153                                node,
2154                                addr >> VMEMMAP_CHUNK_SHIFT,
2155                                VMEMMAP_SIZE);
2156                 }
2157         }
2158         return 0;
2159 }
2160 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
2161
2162 static void prot_init_common(unsigned long page_none,
2163                              unsigned long page_shared,
2164                              unsigned long page_copy,
2165                              unsigned long page_readonly,
2166                              unsigned long page_exec_bit)
2167 {
2168         PAGE_COPY = __pgprot(page_copy);
2169         PAGE_SHARED = __pgprot(page_shared);
2170
2171         protection_map[0x0] = __pgprot(page_none);
2172         protection_map[0x1] = __pgprot(page_readonly & ~page_exec_bit);
2173         protection_map[0x2] = __pgprot(page_copy & ~page_exec_bit);
2174         protection_map[0x3] = __pgprot(page_copy & ~page_exec_bit);
2175         protection_map[0x4] = __pgprot(page_readonly);
2176         protection_map[0x5] = __pgprot(page_readonly);
2177         protection_map[0x6] = __pgprot(page_copy);
2178         protection_map[0x7] = __pgprot(page_copy);
2179         protection_map[0x8] = __pgprot(page_none);
2180         protection_map[0x9] = __pgprot(page_readonly & ~page_exec_bit);
2181         protection_map[0xa] = __pgprot(page_shared & ~page_exec_bit);
2182         protection_map[0xb] = __pgprot(page_shared & ~page_exec_bit);
2183         protection_map[0xc] = __pgprot(page_readonly);
2184         protection_map[0xd] = __pgprot(page_readonly);
2185         protection_map[0xe] = __pgprot(page_shared);
2186         protection_map[0xf] = __pgprot(page_shared);
2187 }
2188
2189 static void __init sun4u_pgprot_init(void)
2190 {
2191         unsigned long page_none, page_shared, page_copy, page_readonly;
2192         unsigned long page_exec_bit;
2193
2194         PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2195                                 _PAGE_CACHE_4U | _PAGE_P_4U |
2196                                 __ACCESS_BITS_4U | __DIRTY_BITS_4U |
2197                                 _PAGE_EXEC_4U);
2198         PAGE_KERNEL_LOCKED = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2199                                        _PAGE_CACHE_4U | _PAGE_P_4U |
2200                                        __ACCESS_BITS_4U | __DIRTY_BITS_4U |
2201                                        _PAGE_EXEC_4U | _PAGE_L_4U);
2202
2203         _PAGE_IE = _PAGE_IE_4U;
2204         _PAGE_E = _PAGE_E_4U;
2205         _PAGE_CACHE = _PAGE_CACHE_4U;
2206
2207         pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4U | __DIRTY_BITS_4U |
2208                      __ACCESS_BITS_4U | _PAGE_E_4U);
2209
2210 #ifdef CONFIG_DEBUG_PAGEALLOC
2211         kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZBITS_4U) ^
2212                 0xfffff80000000000UL;
2213 #else
2214         kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4U) ^
2215                 0xfffff80000000000UL;
2216 #endif
2217         kern_linear_pte_xor[0] |= (_PAGE_CP_4U | _PAGE_CV_4U |
2218                                    _PAGE_P_4U | _PAGE_W_4U);
2219
2220         /* XXX Should use 256MB on Panther. XXX */
2221         kern_linear_pte_xor[1] = kern_linear_pte_xor[0];
2222
2223         _PAGE_SZBITS = _PAGE_SZBITS_4U;
2224         _PAGE_ALL_SZ_BITS =  (_PAGE_SZ4MB_4U | _PAGE_SZ512K_4U |
2225                               _PAGE_SZ64K_4U | _PAGE_SZ8K_4U |
2226                               _PAGE_SZ32MB_4U | _PAGE_SZ256MB_4U);
2227
2228
2229         page_none = _PAGE_PRESENT_4U | _PAGE_ACCESSED_4U | _PAGE_CACHE_4U;
2230         page_shared = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2231                        __ACCESS_BITS_4U | _PAGE_WRITE_4U | _PAGE_EXEC_4U);
2232         page_copy   = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2233                        __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2234         page_readonly   = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2235                            __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2236
2237         page_exec_bit = _PAGE_EXEC_4U;
2238
2239         prot_init_common(page_none, page_shared, page_copy, page_readonly,
2240                          page_exec_bit);
2241 }
2242
2243 static void __init sun4v_pgprot_init(void)
2244 {
2245         unsigned long page_none, page_shared, page_copy, page_readonly;
2246         unsigned long page_exec_bit;
2247
2248         PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4V | _PAGE_VALID |
2249                                 _PAGE_CACHE_4V | _PAGE_P_4V |
2250                                 __ACCESS_BITS_4V | __DIRTY_BITS_4V |
2251                                 _PAGE_EXEC_4V);
2252         PAGE_KERNEL_LOCKED = PAGE_KERNEL;
2253
2254         _PAGE_IE = _PAGE_IE_4V;
2255         _PAGE_E = _PAGE_E_4V;
2256         _PAGE_CACHE = _PAGE_CACHE_4V;
2257
2258 #ifdef CONFIG_DEBUG_PAGEALLOC
2259         kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZBITS_4V) ^
2260                 0xfffff80000000000UL;
2261 #else
2262         kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4V) ^
2263                 0xfffff80000000000UL;
2264 #endif
2265         kern_linear_pte_xor[0] |= (_PAGE_CP_4V | _PAGE_CV_4V |
2266                                    _PAGE_P_4V | _PAGE_W_4V);
2267
2268 #ifdef CONFIG_DEBUG_PAGEALLOC
2269         kern_linear_pte_xor[1] = (_PAGE_VALID | _PAGE_SZBITS_4V) ^
2270                 0xfffff80000000000UL;
2271 #else
2272         kern_linear_pte_xor[1] = (_PAGE_VALID | _PAGE_SZ256MB_4V) ^
2273                 0xfffff80000000000UL;
2274 #endif
2275         kern_linear_pte_xor[1] |= (_PAGE_CP_4V | _PAGE_CV_4V |
2276                                    _PAGE_P_4V | _PAGE_W_4V);
2277
2278         pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4V | __DIRTY_BITS_4V |
2279                      __ACCESS_BITS_4V | _PAGE_E_4V);
2280
2281         _PAGE_SZBITS = _PAGE_SZBITS_4V;
2282         _PAGE_ALL_SZ_BITS = (_PAGE_SZ16GB_4V | _PAGE_SZ2GB_4V |
2283                              _PAGE_SZ256MB_4V | _PAGE_SZ32MB_4V |
2284                              _PAGE_SZ4MB_4V | _PAGE_SZ512K_4V |
2285                              _PAGE_SZ64K_4V | _PAGE_SZ8K_4V);
2286
2287         page_none = _PAGE_PRESENT_4V | _PAGE_ACCESSED_4V | _PAGE_CACHE_4V;
2288         page_shared = (_PAGE_VALID | _PAGE_PRESENT_4V | _PAGE_CACHE_4V |
2289                        __ACCESS_BITS_4V | _PAGE_WRITE_4V | _PAGE_EXEC_4V);
2290         page_copy   = (_PAGE_VALID | _PAGE_PRESENT_4V | _PAGE_CACHE_4V |
2291                        __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2292         page_readonly = (_PAGE_VALID | _PAGE_PRESENT_4V | _PAGE_CACHE_4V |
2293                          __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2294
2295         page_exec_bit = _PAGE_EXEC_4V;
2296
2297         prot_init_common(page_none, page_shared, page_copy, page_readonly,
2298                          page_exec_bit);
2299 }
2300
2301 unsigned long pte_sz_bits(unsigned long sz)
2302 {
2303         if (tlb_type == hypervisor) {
2304                 switch (sz) {
2305                 case 8 * 1024:
2306                 default:
2307                         return _PAGE_SZ8K_4V;
2308                 case 64 * 1024:
2309                         return _PAGE_SZ64K_4V;
2310                 case 512 * 1024:
2311                         return _PAGE_SZ512K_4V;
2312                 case 4 * 1024 * 1024:
2313                         return _PAGE_SZ4MB_4V;
2314                 }
2315         } else {
2316                 switch (sz) {
2317                 case 8 * 1024:
2318                 default:
2319                         return _PAGE_SZ8K_4U;
2320                 case 64 * 1024:
2321                         return _PAGE_SZ64K_4U;
2322                 case 512 * 1024:
2323                         return _PAGE_SZ512K_4U;
2324                 case 4 * 1024 * 1024:
2325                         return _PAGE_SZ4MB_4U;
2326                 }
2327         }
2328 }
2329
2330 pte_t mk_pte_io(unsigned long page, pgprot_t prot, int space, unsigned long page_size)
2331 {
2332         pte_t pte;
2333
2334         pte_val(pte)  = page | pgprot_val(pgprot_noncached(prot));
2335         pte_val(pte) |= (((unsigned long)space) << 32);
2336         pte_val(pte) |= pte_sz_bits(page_size);
2337
2338         return pte;
2339 }
2340
2341 static unsigned long kern_large_tte(unsigned long paddr)
2342 {
2343         unsigned long val;
2344
2345         val = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2346                _PAGE_CP_4U | _PAGE_CV_4U | _PAGE_P_4U |
2347                _PAGE_EXEC_4U | _PAGE_L_4U | _PAGE_W_4U);
2348         if (tlb_type == hypervisor)
2349                 val = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2350                        _PAGE_CP_4V | _PAGE_CV_4V | _PAGE_P_4V |
2351                        _PAGE_EXEC_4V | _PAGE_W_4V);
2352
2353         return val | paddr;
2354 }
2355
2356 /* If not locked, zap it. */
2357 void __flush_tlb_all(void)
2358 {
2359         unsigned long pstate;
2360         int i;
2361
2362         __asm__ __volatile__("flushw\n\t"
2363                              "rdpr      %%pstate, %0\n\t"
2364                              "wrpr      %0, %1, %%pstate"
2365                              : "=r" (pstate)
2366                              : "i" (PSTATE_IE));
2367         if (tlb_type == hypervisor) {
2368                 sun4v_mmu_demap_all();
2369         } else if (tlb_type == spitfire) {
2370                 for (i = 0; i < 64; i++) {
2371                         /* Spitfire Errata #32 workaround */
2372                         /* NOTE: Always runs on spitfire, so no
2373                          *       cheetah+ page size encodings.
2374                          */
2375                         __asm__ __volatile__("stxa      %0, [%1] %2\n\t"
2376                                              "flush     %%g6"
2377                                              : /* No outputs */
2378                                              : "r" (0),
2379                                              "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2380
2381                         if (!(spitfire_get_dtlb_data(i) & _PAGE_L_4U)) {
2382                                 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2383                                                      "membar #Sync"
2384                                                      : /* no outputs */
2385                                                      : "r" (TLB_TAG_ACCESS), "i" (ASI_DMMU));
2386                                 spitfire_put_dtlb_data(i, 0x0UL);
2387                         }
2388
2389                         /* Spitfire Errata #32 workaround */
2390                         /* NOTE: Always runs on spitfire, so no
2391                          *       cheetah+ page size encodings.
2392                          */
2393                         __asm__ __volatile__("stxa      %0, [%1] %2\n\t"
2394                                              "flush     %%g6"
2395                                              : /* No outputs */
2396                                              : "r" (0),
2397                                              "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2398
2399                         if (!(spitfire_get_itlb_data(i) & _PAGE_L_4U)) {
2400                                 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2401                                                      "membar #Sync"
2402                                                      : /* no outputs */
2403                                                      : "r" (TLB_TAG_ACCESS), "i" (ASI_IMMU));
2404                                 spitfire_put_itlb_data(i, 0x0UL);
2405                         }
2406                 }
2407         } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
2408                 cheetah_flush_dtlb_all();
2409                 cheetah_flush_itlb_all();
2410         }
2411         __asm__ __volatile__("wrpr      %0, 0, %%pstate"
2412                              : : "r" (pstate));
2413 }