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/ioport.h>
  26 #include <linux/percpu.h>
  27 #include <linux/memblock.h>
  28 #include <linux/mmzone.h>
  29 #include <linux/gfp.h>
  30
  31 #include <asm/head.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/setup.h>
  52 #include <asm/irq.h>
  53
  54 #include "init_64.h"
  55
  56 unsigned long kern_linear_pte_xor[4] __read_mostly;
  57 static unsigned long page_cache4v_flag;
  58
  59 /* A bitmap, two bits for every 256MB of physical memory.  These two
  60  * bits determine what page size we use for kernel linear
  61  * translations.  They form an index into kern_linear_pte_xor[].  The
  62  * value in the indexed slot is XOR'd with the TLB miss virtual
  63  * address to form the resulting TTE.  The mapping is:
  64  *
  65  *      0       ==>     4MB
  66  *      1       ==>     256MB
  67  *      2       ==>     2GB
  68  *      3       ==>     16GB
  69  *
  70  * All sun4v chips support 256MB pages.  Only SPARC-T4 and later
  71  * support 2GB pages, and hopefully future cpus will support the 16GB
  72  * pages as well.  For slots 2 and 3, we encode a 256MB TTE xor there
  73  * if these larger page sizes are not supported by the cpu.
  74  *
  75  * It would be nice to determine this from the machine description
  76  * 'cpu' properties, but we need to have this table setup before the
  77  * MDESC is initialized.
  78  */
  79
  80 #ifndef CONFIG_DEBUG_PAGEALLOC
  81 /* A special kernel TSB for 4MB, 256MB, 2GB and 16GB linear mappings.
  82  * Space is allocated for this right after the trap table in
  83  * arch/sparc64/kernel/head.S
  84  */
  85 extern struct tsb swapper_4m_tsb[KERNEL_TSB4M_NENTRIES];
  86 #endif
  87 extern struct tsb swapper_tsb[KERNEL_TSB_NENTRIES];
  88
  89 static unsigned long cpu_pgsz_mask;
  90
  91 #define MAX_BANKS       1024
  92
  93 static struct linux_prom64_registers pavail[MAX_BANKS];
  94 static int pavail_ents;
  95
  96 u64 numa_latency[MAX_NUMNODES][MAX_NUMNODES];
  97
  98 static int cmp_p64(const void *a, const void *b)
  99 {
 100         const struct linux_prom64_registers *x = a, *y = b;
 101
 102         if (x->phys_addr > y->phys_addr)
 103                 return 1;
 104         if (x->phys_addr < y->phys_addr)
 105                 return -1;
 106         return 0;
 107 }
 108
 109 static void __init read_obp_memory(const char *property,
 110                                    struct linux_prom64_registers *regs,
 111                                    int *num_ents)
 112 {
 113         phandle node = prom_finddevice("/memory");
 114         int prop_size = prom_getproplen(node, property);
 115         int ents, ret, i;
 116
 117         ents = prop_size / sizeof(struct linux_prom64_registers);
 118         if (ents > MAX_BANKS) {
 119                 prom_printf("The machine has more %s property entries than "
 120                             "this kernel can support (%d).\n",
 121                             property, MAX_BANKS);
 122                 prom_halt();
 123         }
 124
 125         ret = prom_getproperty(node, property, (char *) regs, prop_size);
 126         if (ret == -1) {
 127                 prom_printf("Couldn't get %s property from /memory.\n",
 128                                 property);
 129                 prom_halt();
 130         }
 131
 132         /* Sanitize what we got from the firmware, by page aligning
 133          * everything.
 134          */
 135         for (i = 0; i < ents; i++) {
 136                 unsigned long base, size;
 137
 138                 base = regs[i].phys_addr;
 139                 size = regs[i].reg_size;
 140
 141                 size &= PAGE_MASK;
 142                 if (base & ~PAGE_MASK) {
 143                         unsigned long new_base = PAGE_ALIGN(base);
 144
 145                         size -= new_base - base;
 146                         if ((long) size < 0L)
 147                                 size = 0UL;
 148                         base = new_base;
 149                 }
 150                 if (size == 0UL) {
 151                         /* If it is empty, simply get rid of it.
 152                          * This simplifies the logic of the other
 153                          * functions that process these arrays.
 154                          */
 155                         memmove(&regs[i], &regs[i + 1],
 156                                 (ents - i - 1) * sizeof(regs[0]));
 157                         i--;
 158                         ents--;
 159                         continue;
 160                 }
 161                 regs[i].phys_addr = base;
 162                 regs[i].reg_size = size;
 163         }
 164
 165         *num_ents = ents;
 166
 167         sort(regs, ents, sizeof(struct linux_prom64_registers),
 168              cmp_p64, NULL);
 169 }
 170
 171 /* Kernel physical address base and size in bytes.  */
 172 unsigned long kern_base __read_mostly;
 173 unsigned long kern_size __read_mostly;
 174
 175 /* Initial ramdisk setup */
 176 extern unsigned long sparc_ramdisk_image64;
 177 extern unsigned int sparc_ramdisk_image;
 178 extern unsigned int sparc_ramdisk_size;
 179
 180 struct page *mem_map_zero __read_mostly;
 181 EXPORT_SYMBOL(mem_map_zero);
 182
 183 unsigned int sparc64_highest_unlocked_tlb_ent __read_mostly;
 184
 185 unsigned long sparc64_kern_pri_context __read_mostly;
 186 unsigned long sparc64_kern_pri_nuc_bits __read_mostly;
 187 unsigned long sparc64_kern_sec_context __read_mostly;
 188
 189 int num_kernel_image_mappings;
 190
 191 #ifdef CONFIG_DEBUG_DCFLUSH
 192 atomic_t dcpage_flushes = ATOMIC_INIT(0);
 193 #ifdef CONFIG_SMP
 194 atomic_t dcpage_flushes_xcall = ATOMIC_INIT(0);
 195 #endif
 196 #endif
 197
 198 inline void flush_dcache_page_impl(struct page *page)
 199 {
 200         BUG_ON(tlb_type == hypervisor);
 201 #ifdef CONFIG_DEBUG_DCFLUSH
 202         atomic_inc(&dcpage_flushes);
 203 #endif
 204
 205 #ifdef DCACHE_ALIASING_POSSIBLE
 206         __flush_dcache_page(page_address(page),
 207                             ((tlb_type == spitfire) &&
 208                              page_mapping(page) != NULL));
 209 #else
 210         if (page_mapping(page) != NULL &&
 211             tlb_type == spitfire)
 212                 __flush_icache_page(__pa(page_address(page)));
 213 #endif
 214 }
 215
 216 #define PG_dcache_dirty         PG_arch_1
 217 #define PG_dcache_cpu_shift     32UL
 218 #define PG_dcache_cpu_mask      \
 219         ((1UL<<ilog2(roundup_pow_of_two(NR_CPUS)))-1UL)
 220
 221 #define dcache_dirty_cpu(page) \
 222         (((page)->flags >> PG_dcache_cpu_shift) & PG_dcache_cpu_mask)
 223
 224 static inline void set_dcache_dirty(struct page *page, int this_cpu)
 225 {
 226         unsigned long mask = this_cpu;
 227         unsigned long non_cpu_bits;
 228
 229         non_cpu_bits = ~(PG_dcache_cpu_mask << PG_dcache_cpu_shift);
 230         mask = (mask << PG_dcache_cpu_shift) | (1UL << PG_dcache_dirty);
 231
 232         __asm__ __volatile__("1:\n\t"
 233                              "ldx       [%2], %%g7\n\t"
 234                              "and       %%g7, %1, %%g1\n\t"
 235                              "or        %%g1, %0, %%g1\n\t"
 236                              "casx      [%2], %%g7, %%g1\n\t"
 237                              "cmp       %%g7, %%g1\n\t"
 238                              "bne,pn    %%xcc, 1b\n\t"
 239                              " nop"
 240                              : /* no outputs */
 241                              : "r" (mask), "r" (non_cpu_bits), "r" (&page->flags)
 242                              : "g1", "g7");
 243 }
 244
 245 static inline void clear_dcache_dirty_cpu(struct page *page, unsigned long cpu)
 246 {
 247         unsigned long mask = (1UL << PG_dcache_dirty);
 248
 249         __asm__ __volatile__("! test_and_clear_dcache_dirty\n"
 250                              "1:\n\t"
 251                              "ldx       [%2], %%g7\n\t"
 252                              "srlx      %%g7, %4, %%g1\n\t"
 253                              "and       %%g1, %3, %%g1\n\t"
 254                              "cmp       %%g1, %0\n\t"
 255                              "bne,pn    %%icc, 2f\n\t"
 256                              " andn     %%g7, %1, %%g1\n\t"
 257                              "casx      [%2], %%g7, %%g1\n\t"
 258                              "cmp       %%g7, %%g1\n\t"
 259                              "bne,pn    %%xcc, 1b\n\t"
 260                              " nop\n"
 261                              "2:"
 262                              : /* no outputs */
 263                              : "r" (cpu), "r" (mask), "r" (&page->flags),
 264                                "i" (PG_dcache_cpu_mask),
 265                                "i" (PG_dcache_cpu_shift)
 266                              : "g1", "g7");
 267 }
 268
 269 static inline void tsb_insert(struct tsb *ent, unsigned long tag, unsigned long pte)
 270 {
 271         unsigned long tsb_addr = (unsigned long) ent;
 272
 273         if (tlb_type == cheetah_plus || tlb_type == hypervisor)
 274                 tsb_addr = __pa(tsb_addr);
 275
 276         __tsb_insert(tsb_addr, tag, pte);
 277 }
 278
 279 unsigned long _PAGE_ALL_SZ_BITS __read_mostly;
 280
 281 static void flush_dcache(unsigned long pfn)
 282 {
 283         struct page *page;
 284
 285         page = pfn_to_page(pfn);
 286         if (page) {
 287                 unsigned long pg_flags;
 288
 289                 pg_flags = page->flags;
 290                 if (pg_flags & (1UL << PG_dcache_dirty)) {
 291                         int cpu = ((pg_flags >> PG_dcache_cpu_shift) &
 292                                    PG_dcache_cpu_mask);
 293                         int this_cpu = get_cpu();
 294
 295                         /* This is just to optimize away some function calls
 296                          * in the SMP case.
 297                          */
 298                         if (cpu == this_cpu)
 299                                 flush_dcache_page_impl(page);
 300                         else
 301                                 smp_flush_dcache_page_impl(page, cpu);
 302
 303                         clear_dcache_dirty_cpu(page, cpu);
 304
 305                         put_cpu();
 306                 }
 307         }
 308 }
 309
 310 /* mm->context.lock must be held */
 311 static void __update_mmu_tsb_insert(struct mm_struct *mm, unsigned long tsb_index,
 312                                     unsigned long tsb_hash_shift, unsigned long address,
 313                                     unsigned long tte)
 314 {
 315         struct tsb *tsb = mm->context.tsb_block[tsb_index].tsb;
 316         unsigned long tag;
 317
 318         if (unlikely(!tsb))
 319                 return;
 320
 321         tsb += ((address >> tsb_hash_shift) &
 322                 (mm->context.tsb_block[tsb_index].tsb_nentries - 1UL));
 323         tag = (address >> 22UL);
 324         tsb_insert(tsb, tag, tte);
 325 }
 326
 327 void update_mmu_cache(struct vm_area_struct *vma, unsigned long address, pte_t *ptep)
 328 {
 329         struct mm_struct *mm;
 330         unsigned long flags;
 331         pte_t pte = *ptep;
 332
 333         if (tlb_type != hypervisor) {
 334                 unsigned long pfn = pte_pfn(pte);
 335
 336                 if (pfn_valid(pfn))
 337                         flush_dcache(pfn);
 338         }
 339
 340         mm = vma->vm_mm;
 341
 342         /* Don't insert a non-valid PTE into the TSB, we'll deadlock.  */
 343         if (!pte_accessible(mm, pte))
 344                 return;
 345
 346         spin_lock_irqsave(&mm->context.lock, flags);
 347
 348 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
 349         if (mm->context.huge_pte_count && is_hugetlb_pte(pte))
 350                 __update_mmu_tsb_insert(mm, MM_TSB_HUGE, REAL_HPAGE_SHIFT,
 351                                         address, pte_val(pte));
 352         else
 353 #endif
 354                 __update_mmu_tsb_insert(mm, MM_TSB_BASE, PAGE_SHIFT,
 355                                         address, pte_val(pte));
 356
 357         spin_unlock_irqrestore(&mm->context.lock, flags);
 358 }
 359
 360 void flush_dcache_page(struct page *page)
 361 {
 362         struct address_space *mapping;
 363         int this_cpu;
 364
 365         if (tlb_type == hypervisor)
 366                 return;
 367
 368         /* Do not bother with the expensive D-cache flush if it
 369          * is merely the zero page.  The 'bigcore' testcase in GDB
 370          * causes this case to run millions of times.
 371          */
 372         if (page == ZERO_PAGE(0))
 373                 return;
 374
 375         this_cpu = get_cpu();
 376
 377         mapping = page_mapping(page);
 378         if (mapping && !mapping_mapped(mapping)) {
 379                 int dirty = test_bit(PG_dcache_dirty, &page->flags);
 380                 if (dirty) {
 381                         int dirty_cpu = dcache_dirty_cpu(page);
 382
 383                         if (dirty_cpu == this_cpu)
 384                                 goto out;
 385                         smp_flush_dcache_page_impl(page, dirty_cpu);
 386                 }
 387                 set_dcache_dirty(page, this_cpu);
 388         } else {
 389                 /* We could delay the flush for the !page_mapping
 390                  * case too.  But that case is for exec env/arg
 391                  * pages and those are %99 certainly going to get
 392                  * faulted into the tlb (and thus flushed) anyways.
 393                  */
 394                 flush_dcache_page_impl(page);
 395         }
 396
 397 out:
 398         put_cpu();
 399 }
 400 EXPORT_SYMBOL(flush_dcache_page);
 401
 402 void __kprobes flush_icache_range(unsigned long start, unsigned long end)
 403 {
 404         /* Cheetah and Hypervisor platform cpus have coherent I-cache. */
 405         if (tlb_type == spitfire) {
 406                 unsigned long kaddr;
 407
 408                 /* This code only runs on Spitfire cpus so this is
 409                  * why we can assume _PAGE_PADDR_4U.
 410                  */
 411                 for (kaddr = start; kaddr < end; kaddr += PAGE_SIZE) {
 412                         unsigned long paddr, mask = _PAGE_PADDR_4U;
 413
 414                         if (kaddr >= PAGE_OFFSET)
 415                                 paddr = kaddr & mask;
 416                         else {
 417                                 pgd_t *pgdp = pgd_offset_k(kaddr);
 418                                 pud_t *pudp = pud_offset(pgdp, kaddr);
 419                                 pmd_t *pmdp = pmd_offset(pudp, kaddr);
 420                                 pte_t *ptep = pte_offset_kernel(pmdp, kaddr);
 421
 422                                 paddr = pte_val(*ptep) & mask;
 423                         }
 424                         __flush_icache_page(paddr);
 425                 }
 426         }
 427 }
 428 EXPORT_SYMBOL(flush_icache_range);
 429
 430 void mmu_info(struct seq_file *m)
 431 {
 432         static const char *pgsz_strings[] = {
 433                 "8K", "64K", "512K", "4MB", "32MB",
 434                 "256MB", "2GB", "16GB",
 435         };
 436         int i, printed;
 437
 438         if (tlb_type == cheetah)
 439                 seq_printf(m, "MMU Type\t: Cheetah\n");
 440         else if (tlb_type == cheetah_plus)
 441                 seq_printf(m, "MMU Type\t: Cheetah+\n");
 442         else if (tlb_type == spitfire)
 443                 seq_printf(m, "MMU Type\t: Spitfire\n");
 444         else if (tlb_type == hypervisor)
 445                 seq_printf(m, "MMU Type\t: Hypervisor (sun4v)\n");
 446         else
 447                 seq_printf(m, "MMU Type\t: ???\n");
 448
 449         seq_printf(m, "MMU PGSZs\t: ");
 450         printed = 0;
 451         for (i = 0; i < ARRAY_SIZE(pgsz_strings); i++) {
 452                 if (cpu_pgsz_mask & (1UL << i)) {
 453                         seq_printf(m, "%s%s",
 454                                    printed ? "," : "", pgsz_strings[i]);
 455                         printed++;
 456                 }
 457         }
 458         seq_putc(m, '\n');
 459
 460 #ifdef CONFIG_DEBUG_DCFLUSH
 461         seq_printf(m, "DCPageFlushes\t: %d\n",
 462                    atomic_read(&dcpage_flushes));
 463 #ifdef CONFIG_SMP
 464         seq_printf(m, "DCPageFlushesXC\t: %d\n",
 465                    atomic_read(&dcpage_flushes_xcall));
 466 #endif /* CONFIG_SMP */
 467 #endif /* CONFIG_DEBUG_DCFLUSH */
 468 }
 469
 470 struct linux_prom_translation prom_trans[512] __read_mostly;
 471 unsigned int prom_trans_ents __read_mostly;
 472
 473 unsigned long kern_locked_tte_data;
 474
 475 /* The obp translations are saved based on 8k pagesize, since obp can
 476  * use a mixture of pagesizes. Misses to the LOW_OBP_ADDRESS ->
 477  * HI_OBP_ADDRESS range are handled in ktlb.S.
 478  */
 479 static inline int in_obp_range(unsigned long vaddr)
 480 {
 481         return (vaddr >= LOW_OBP_ADDRESS &&
 482                 vaddr < HI_OBP_ADDRESS);
 483 }
 484
 485 static int cmp_ptrans(const void *a, const void *b)
 486 {
 487         const struct linux_prom_translation *x = a, *y = b;
 488
 489         if (x->virt > y->virt)
 490                 return 1;
 491         if (x->virt < y->virt)
 492                 return -1;
 493         return 0;
 494 }
 495
 496 /* Read OBP translations property into 'prom_trans[]'.  */
 497 static void __init read_obp_translations(void)
 498 {
 499         int n, node, ents, first, last, i;
 500
 501         node = prom_finddevice("/virtual-memory");
 502         n = prom_getproplen(node, "translations");
 503         if (unlikely(n == 0 || n == -1)) {
 504                 prom_printf("prom_mappings: Couldn't get size.\n");
 505                 prom_halt();
 506         }
 507         if (unlikely(n > sizeof(prom_trans))) {
 508                 prom_printf("prom_mappings: Size %d is too big.\n", n);
 509                 prom_halt();
 510         }
 511
 512         if ((n = prom_getproperty(node, "translations",
 513                                   (char *)&prom_trans[0],
 514                                   sizeof(prom_trans))) == -1) {
 515                 prom_printf("prom_mappings: Couldn't get property.\n");
 516                 prom_halt();
 517         }
 518
 519         n = n / sizeof(struct linux_prom_translation);
 520
 521         ents = n;
 522
 523         sort(prom_trans, ents, sizeof(struct linux_prom_translation),
 524              cmp_ptrans, NULL);
 525
 526         /* Now kick out all the non-OBP entries.  */
 527         for (i = 0; i < ents; i++) {
 528                 if (in_obp_range(prom_trans[i].virt))
 529                         break;
 530         }
 531         first = i;
 532         for (; i < ents; i++) {
 533                 if (!in_obp_range(prom_trans[i].virt))
 534                         break;
 535         }
 536         last = i;
 537
 538         for (i = 0; i < (last - first); i++) {
 539                 struct linux_prom_translation *src = &prom_trans[i + first];
 540                 struct linux_prom_translation *dest = &prom_trans[i];
 541
 542                 *dest = *src;
 543         }
 544         for (; i < ents; i++) {
 545                 struct linux_prom_translation *dest = &prom_trans[i];
 546                 dest->virt = dest->size = dest->data = 0x0UL;
 547         }
 548
 549         prom_trans_ents = last - first;
 550
 551         if (tlb_type == spitfire) {
 552                 /* Clear diag TTE bits. */
 553                 for (i = 0; i < prom_trans_ents; i++)
 554                         prom_trans[i].data &= ~0x0003fe0000000000UL;
 555         }
 556
 557         /* Force execute bit on.  */
 558         for (i = 0; i < prom_trans_ents; i++)
 559                 prom_trans[i].data |= (tlb_type == hypervisor ?
 560                                        _PAGE_EXEC_4V : _PAGE_EXEC_4U);
 561 }
 562
 563 static void __init hypervisor_tlb_lock(unsigned long vaddr,
 564                                        unsigned long pte,
 565                                        unsigned long mmu)
 566 {
 567         unsigned long ret = sun4v_mmu_map_perm_addr(vaddr, 0, pte, mmu);
 568
 569         if (ret != 0) {
 570                 prom_printf("hypervisor_tlb_lock[%lx:%x:%lx:%lx]: "
 571                             "errors with %lx\n", vaddr, 0, pte, mmu, ret);
 572                 prom_halt();
 573         }
 574 }
 575
 576 static unsigned long kern_large_tte(unsigned long paddr);
 577
 578 static void __init remap_kernel(void)
 579 {
 580         unsigned long phys_page, tte_vaddr, tte_data;
 581         int i, tlb_ent = sparc64_highest_locked_tlbent();
 582
 583         tte_vaddr = (unsigned long) KERNBASE;
 584         phys_page = (prom_boot_mapping_phys_low >> ILOG2_4MB) << ILOG2_4MB;
 585         tte_data = kern_large_tte(phys_page);
 586
 587         kern_locked_tte_data = tte_data;
 588
 589         /* Now lock us into the TLBs via Hypervisor or OBP. */
 590         if (tlb_type == hypervisor) {
 591                 for (i = 0; i < num_kernel_image_mappings; i++) {
 592                         hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_DMMU);
 593                         hypervisor_tlb_lock(tte_vaddr, tte_data, HV_MMU_IMMU);
 594                         tte_vaddr += 0x400000;
 595                         tte_data += 0x400000;
 596                 }
 597         } else {
 598                 for (i = 0; i < num_kernel_image_mappings; i++) {
 599                         prom_dtlb_load(tlb_ent - i, tte_data, tte_vaddr);
 600                         prom_itlb_load(tlb_ent - i, tte_data, tte_vaddr);
 601                         tte_vaddr += 0x400000;
 602                         tte_data += 0x400000;
 603                 }
 604                 sparc64_highest_unlocked_tlb_ent = tlb_ent - i;
 605         }
 606         if (tlb_type == cheetah_plus) {
 607                 sparc64_kern_pri_context = (CTX_CHEETAH_PLUS_CTX0 |
 608                                             CTX_CHEETAH_PLUS_NUC);
 609                 sparc64_kern_pri_nuc_bits = CTX_CHEETAH_PLUS_NUC;
 610                 sparc64_kern_sec_context = CTX_CHEETAH_PLUS_CTX0;
 611         }
 612 }
 613
 614
 615 static void __init inherit_prom_mappings(void)
 616 {
 617         /* Now fixup OBP's idea about where we really are mapped. */
 618         printk("Remapping the kernel... ");
 619         remap_kernel();
 620         printk("done.\n");
 621 }
 622
 623 void prom_world(int enter)
 624 {
 625         if (!enter)
 626                 set_fs(get_fs());
 627
 628         __asm__ __volatile__("flushw");
 629 }
 630
 631 void __flush_dcache_range(unsigned long start, unsigned long end)
 632 {
 633         unsigned long va;
 634
 635         if (tlb_type == spitfire) {
 636                 int n = 0;
 637
 638                 for (va = start; va < end; va += 32) {
 639                         spitfire_put_dcache_tag(va & 0x3fe0, 0x0);
 640                         if (++n >= 512)
 641                                 break;
 642                 }
 643         } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
 644                 start = __pa(start);
 645                 end = __pa(end);
 646                 for (va = start; va < end; va += 32)
 647                         __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
 648                                              "membar #Sync"
 649                                              : /* no outputs */
 650                                              : "r" (va),
 651                                                "i" (ASI_DCACHE_INVALIDATE));
 652         }
 653 }
 654 EXPORT_SYMBOL(__flush_dcache_range);
 655
 656 /* get_new_mmu_context() uses "cache + 1".  */
 657 DEFINE_SPINLOCK(ctx_alloc_lock);
 658 unsigned long tlb_context_cache = CTX_FIRST_VERSION - 1;
 659 #define MAX_CTX_NR      (1UL << CTX_NR_BITS)
 660 #define CTX_BMAP_SLOTS  BITS_TO_LONGS(MAX_CTX_NR)
 661 DECLARE_BITMAP(mmu_context_bmap, MAX_CTX_NR);
 662
 663 /* Caller does TLB context flushing on local CPU if necessary.
 664  * The caller also ensures that CTX_VALID(mm->context) is false.
 665  *
 666  * We must be careful about boundary cases so that we never
 667  * let the user have CTX 0 (nucleus) or we ever use a CTX
 668  * version of zero (and thus NO_CONTEXT would not be caught
 669  * by version mis-match tests in mmu_context.h).
 670  *
 671  * Always invoked with interrupts disabled.
 672  */
 673 void get_new_mmu_context(struct mm_struct *mm)
 674 {
 675         unsigned long ctx, new_ctx;
 676         unsigned long orig_pgsz_bits;
 677         int new_version;
 678
 679         spin_lock(&ctx_alloc_lock);
 680         orig_pgsz_bits = (mm->context.sparc64_ctx_val & CTX_PGSZ_MASK);
 681         ctx = (tlb_context_cache + 1) & CTX_NR_MASK;
 682         new_ctx = find_next_zero_bit(mmu_context_bmap, 1 << CTX_NR_BITS, ctx);
 683         new_version = 0;
 684         if (new_ctx >= (1 << CTX_NR_BITS)) {
 685                 new_ctx = find_next_zero_bit(mmu_context_bmap, ctx, 1);
 686                 if (new_ctx >= ctx) {
 687                         int i;
 688                         new_ctx = (tlb_context_cache & CTX_VERSION_MASK) +
 689                                 CTX_FIRST_VERSION;
 690                         if (new_ctx == 1)
 691                                 new_ctx = CTX_FIRST_VERSION;
 692
 693                         /* Don't call memset, for 16 entries that's just
 694                          * plain silly...
 695                          */
 696                         mmu_context_bmap[0] = 3;
 697                         mmu_context_bmap[1] = 0;
 698                         mmu_context_bmap[2] = 0;
 699                         mmu_context_bmap[3] = 0;
 700                         for (i = 4; i < CTX_BMAP_SLOTS; i += 4) {
 701                                 mmu_context_bmap[i + 0] = 0;
 702                                 mmu_context_bmap[i + 1] = 0;
 703                                 mmu_context_bmap[i + 2] = 0;
 704                                 mmu_context_bmap[i + 3] = 0;
 705                         }
 706                         new_version = 1;
 707                         goto out;
 708                 }
 709         }
 710         mmu_context_bmap[new_ctx>>6] |= (1UL << (new_ctx & 63));
 711         new_ctx |= (tlb_context_cache & CTX_VERSION_MASK);
 712 out:
 713         tlb_context_cache = new_ctx;
 714         mm->context.sparc64_ctx_val = new_ctx | orig_pgsz_bits;
 715         spin_unlock(&ctx_alloc_lock);
 716
 717         if (unlikely(new_version))
 718                 smp_new_mmu_context_version();
 719 }
 720
 721 static int numa_enabled = 1;
 722 static int numa_debug;
 723
 724 static int __init early_numa(char *p)
 725 {
 726         if (!p)
 727                 return 0;
 728
 729         if (strstr(p, "off"))
 730                 numa_enabled = 0;
 731
 732         if (strstr(p, "debug"))
 733                 numa_debug = 1;
 734
 735         return 0;
 736 }
 737 early_param("numa", early_numa);
 738
 739 #define numadbg(f, a...) \
 740 do {    if (numa_debug) \
 741                 printk(KERN_INFO f, ## a); \
 742 } while (0)
 743
 744 static void __init find_ramdisk(unsigned long phys_base)
 745 {
 746 #ifdef CONFIG_BLK_DEV_INITRD
 747         if (sparc_ramdisk_image || sparc_ramdisk_image64) {
 748                 unsigned long ramdisk_image;
 749
 750                 /* Older versions of the bootloader only supported a
 751                  * 32-bit physical address for the ramdisk image
 752                  * location, stored at sparc_ramdisk_image.  Newer
 753                  * SILO versions set sparc_ramdisk_image to zero and
 754                  * provide a full 64-bit physical address at
 755                  * sparc_ramdisk_image64.
 756                  */
 757                 ramdisk_image = sparc_ramdisk_image;
 758                 if (!ramdisk_image)
 759                         ramdisk_image = sparc_ramdisk_image64;
 760
 761                 /* Another bootloader quirk.  The bootloader normalizes
 762                  * the physical address to KERNBASE, so we have to
 763                  * factor that back out and add in the lowest valid
 764                  * physical page address to get the true physical address.
 765                  */
 766                 ramdisk_image -= KERNBASE;
 767                 ramdisk_image += phys_base;
 768
 769                 numadbg("Found ramdisk at physical address 0x%lx, size %u\n",
 770                         ramdisk_image, sparc_ramdisk_size);
 771
 772                 initrd_start = ramdisk_image;
 773                 initrd_end = ramdisk_image + sparc_ramdisk_size;
 774
 775                 memblock_reserve(initrd_start, sparc_ramdisk_size);
 776
 777                 initrd_start += PAGE_OFFSET;
 778                 initrd_end += PAGE_OFFSET;
 779         }
 780 #endif
 781 }
 782
 783 struct node_mem_mask {
 784         unsigned long mask;
 785         unsigned long val;
 786 };
 787 static struct node_mem_mask node_masks[MAX_NUMNODES];
 788 static int num_node_masks;
 789
 790 #ifdef CONFIG_NEED_MULTIPLE_NODES
 791
 792 int numa_cpu_lookup_table[NR_CPUS];
 793 cpumask_t numa_cpumask_lookup_table[MAX_NUMNODES];
 794
 795 struct mdesc_mblock {
 796         u64     base;
 797         u64     size;
 798         u64     offset; /* RA-to-PA */
 799 };
 800 static struct mdesc_mblock *mblocks;
 801 static int num_mblocks;
 802
 803 static unsigned long ra_to_pa(unsigned long addr)
 804 {
 805         int i;
 806
 807         for (i = 0; i < num_mblocks; i++) {
 808                 struct mdesc_mblock *m = &mblocks[i];
 809
 810                 if (addr >= m->base &&
 811                     addr < (m->base + m->size)) {
 812                         addr += m->offset;
 813                         break;
 814                 }
 815         }
 816         return addr;
 817 }
 818
 819 static int find_node(unsigned long addr)
 820 {
 821         int i;
 822
 823         addr = ra_to_pa(addr);
 824         for (i = 0; i < num_node_masks; i++) {
 825                 struct node_mem_mask *p = &node_masks[i];
 826
 827                 if ((addr & p->mask) == p->val)
 828                         return i;
 829         }
 830         /* The following condition has been observed on LDOM guests.*/
 831         WARN_ONCE(1, "find_node: A physical address doesn't match a NUMA node"
 832                 " rule. Some physical memory will be owned by node 0.");
 833         return 0;
 834 }
 835
 836 static u64 memblock_nid_range(u64 start, u64 end, int *nid)
 837 {
 838         *nid = find_node(start);
 839         start += PAGE_SIZE;
 840         while (start < end) {
 841                 int n = find_node(start);
 842
 843                 if (n != *nid)
 844                         break;
 845                 start += PAGE_SIZE;
 846         }
 847
 848         if (start > end)
 849                 start = end;
 850
 851         return start;
 852 }
 853 #endif
 854
 855 /* This must be invoked after performing all of the necessary
 856  * memblock_set_node() calls for 'nid'.  We need to be able to get
 857  * correct data from get_pfn_range_for_nid().
 858  */
 859 static void __init allocate_node_data(int nid)
 860 {
 861         struct pglist_data *p;
 862         unsigned long start_pfn, end_pfn;
 863 #ifdef CONFIG_NEED_MULTIPLE_NODES
 864         unsigned long paddr;
 865
 866         paddr = memblock_alloc_try_nid(sizeof(struct pglist_data), SMP_CACHE_BYTES, nid);
 867         if (!paddr) {
 868                 prom_printf("Cannot allocate pglist_data for nid[%d]\n", nid);
 869                 prom_halt();
 870         }
 871         NODE_DATA(nid) = __va(paddr);
 872         memset(NODE_DATA(nid), 0, sizeof(struct pglist_data));
 873
 874         NODE_DATA(nid)->node_id = nid;
 875 #endif
 876
 877         p = NODE_DATA(nid);
 878
 879         get_pfn_range_for_nid(nid, &start_pfn, &end_pfn);
 880         p->node_start_pfn = start_pfn;
 881         p->node_spanned_pages = end_pfn - start_pfn;
 882 }
 883
 884 static void init_node_masks_nonnuma(void)
 885 {
 886 #ifdef CONFIG_NEED_MULTIPLE_NODES
 887         int i;
 888 #endif
 889
 890         numadbg("Initializing tables for non-numa.\n");
 891
 892         node_masks[0].mask = node_masks[0].val = 0;
 893         num_node_masks = 1;
 894
 895 #ifdef CONFIG_NEED_MULTIPLE_NODES
 896         for (i = 0; i < NR_CPUS; i++)
 897                 numa_cpu_lookup_table[i] = 0;
 898
 899         cpumask_setall(&numa_cpumask_lookup_table[0]);
 900 #endif
 901 }
 902
 903 #ifdef CONFIG_NEED_MULTIPLE_NODES
 904 struct pglist_data *node_data[MAX_NUMNODES];
 905
 906 EXPORT_SYMBOL(numa_cpu_lookup_table);
 907 EXPORT_SYMBOL(numa_cpumask_lookup_table);
 908 EXPORT_SYMBOL(node_data);
 909
 910 struct mdesc_mlgroup {
 911         u64     node;
 912         u64     latency;
 913         u64     match;
 914         u64     mask;
 915 };
 916 static struct mdesc_mlgroup *mlgroups;
 917 static int num_mlgroups;
 918
 919 static int scan_pio_for_cfg_handle(struct mdesc_handle *md, u64 pio,
 920                                    u32 cfg_handle)
 921 {
 922         u64 arc;
 923
 924         mdesc_for_each_arc(arc, md, pio, MDESC_ARC_TYPE_FWD) {
 925                 u64 target = mdesc_arc_target(md, arc);
 926                 const u64 *val;
 927
 928                 val = mdesc_get_property(md, target,
 929                                          "cfg-handle", NULL);
 930                 if (val && *val == cfg_handle)
 931                         return 0;
 932         }
 933         return -ENODEV;
 934 }
 935
 936 static int scan_arcs_for_cfg_handle(struct mdesc_handle *md, u64 grp,
 937                                     u32 cfg_handle)
 938 {
 939         u64 arc, candidate, best_latency = ~(u64)0;
 940
 941         candidate = MDESC_NODE_NULL;
 942         mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
 943                 u64 target = mdesc_arc_target(md, arc);
 944                 const char *name = mdesc_node_name(md, target);
 945                 const u64 *val;
 946
 947                 if (strcmp(name, "pio-latency-group"))
 948                         continue;
 949
 950                 val = mdesc_get_property(md, target, "latency", NULL);
 951                 if (!val)
 952                         continue;
 953
 954                 if (*val < best_latency) {
 955                         candidate = target;
 956                         best_latency = *val;
 957                 }
 958         }
 959
 960         if (candidate == MDESC_NODE_NULL)
 961                 return -ENODEV;
 962
 963         return scan_pio_for_cfg_handle(md, candidate, cfg_handle);
 964 }
 965
 966 int of_node_to_nid(struct device_node *dp)
 967 {
 968         const struct linux_prom64_registers *regs;
 969         struct mdesc_handle *md;
 970         u32 cfg_handle;
 971         int count, nid;
 972         u64 grp;
 973
 974         /* This is the right thing to do on currently supported
 975          * SUN4U NUMA platforms as well, as the PCI controller does
 976          * not sit behind any particular memory controller.
 977          */
 978         if (!mlgroups)
 979                 return -1;
 980
 981         regs = of_get_property(dp, "reg", NULL);
 982         if (!regs)
 983                 return -1;
 984
 985         cfg_handle = (regs->phys_addr >> 32UL) & 0x0fffffff;
 986
 987         md = mdesc_grab();
 988
 989         count = 0;
 990         nid = -1;
 991         mdesc_for_each_node_by_name(md, grp, "group") {
 992                 if (!scan_arcs_for_cfg_handle(md, grp, cfg_handle)) {
 993                         nid = count;
 994                         break;
 995                 }
 996                 count++;
 997         }
 998
 999         mdesc_release(md);
1000
1001         return nid;
1002 }
1003
1004 static void __init add_node_ranges(void)
1005 {
1006         struct memblock_region *reg;
1007
1008         for_each_memblock(memory, reg) {
1009                 unsigned long size = reg->size;
1010                 unsigned long start, end;
1011
1012                 start = reg->base;
1013                 end = start + size;
1014                 while (start < end) {
1015                         unsigned long this_end;
1016                         int nid;
1017
1018                         this_end = memblock_nid_range(start, end, &nid);
1019
1020                         numadbg("Setting memblock NUMA node nid[%d] "
1021                                 "start[%lx] end[%lx]\n",
1022                                 nid, start, this_end);
1023
1024                         memblock_set_node(start, this_end - start,
1025                                           &memblock.memory, nid);
1026                         start = this_end;
1027                 }
1028         }
1029 }
1030
1031 static int __init grab_mlgroups(struct mdesc_handle *md)
1032 {
1033         unsigned long paddr;
1034         int count = 0;
1035         u64 node;
1036
1037         mdesc_for_each_node_by_name(md, node, "memory-latency-group")
1038                 count++;
1039         if (!count)
1040                 return -ENOENT;
1041
1042         paddr = memblock_alloc(count * sizeof(struct mdesc_mlgroup),
1043                           SMP_CACHE_BYTES);
1044         if (!paddr)
1045                 return -ENOMEM;
1046
1047         mlgroups = __va(paddr);
1048         num_mlgroups = count;
1049
1050         count = 0;
1051         mdesc_for_each_node_by_name(md, node, "memory-latency-group") {
1052                 struct mdesc_mlgroup *m = &mlgroups[count++];
1053                 const u64 *val;
1054
1055                 m->node = node;
1056
1057                 val = mdesc_get_property(md, node, "latency", NULL);
1058                 m->latency = *val;
1059                 val = mdesc_get_property(md, node, "address-match", NULL);
1060                 m->match = *val;
1061                 val = mdesc_get_property(md, node, "address-mask", NULL);
1062                 m->mask = *val;
1063
1064                 numadbg("MLGROUP[%d]: node[%llx] latency[%llx] "
1065                         "match[%llx] mask[%llx]\n",
1066                         count - 1, m->node, m->latency, m->match, m->mask);
1067         }
1068
1069         return 0;
1070 }
1071
1072 static int __init grab_mblocks(struct mdesc_handle *md)
1073 {
1074         unsigned long paddr;
1075         int count = 0;
1076         u64 node;
1077
1078         mdesc_for_each_node_by_name(md, node, "mblock")
1079                 count++;
1080         if (!count)
1081                 return -ENOENT;
1082
1083         paddr = memblock_alloc(count * sizeof(struct mdesc_mblock),
1084                           SMP_CACHE_BYTES);
1085         if (!paddr)
1086                 return -ENOMEM;
1087
1088         mblocks = __va(paddr);
1089         num_mblocks = count;
1090
1091         count = 0;
1092         mdesc_for_each_node_by_name(md, node, "mblock") {
1093                 struct mdesc_mblock *m = &mblocks[count++];
1094                 const u64 *val;
1095
1096                 val = mdesc_get_property(md, node, "base", NULL);
1097                 m->base = *val;
1098                 val = mdesc_get_property(md, node, "size", NULL);
1099                 m->size = *val;
1100                 val = mdesc_get_property(md, node,
1101                                          "address-congruence-offset", NULL);
1102
1103                 /* The address-congruence-offset property is optional.
1104                  * Explicity zero it be identifty this.
1105                  */
1106                 if (val)
1107                         m->offset = *val;
1108                 else
1109                         m->offset = 0UL;
1110
1111                 numadbg("MBLOCK[%d]: base[%llx] size[%llx] offset[%llx]\n",
1112                         count - 1, m->base, m->size, m->offset);
1113         }
1114
1115         return 0;
1116 }
1117
1118 static void __init numa_parse_mdesc_group_cpus(struct mdesc_handle *md,
1119                                                u64 grp, cpumask_t *mask)
1120 {
1121         u64 arc;
1122
1123         cpumask_clear(mask);
1124
1125         mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_BACK) {
1126                 u64 target = mdesc_arc_target(md, arc);
1127                 const char *name = mdesc_node_name(md, target);
1128                 const u64 *id;
1129
1130                 if (strcmp(name, "cpu"))
1131                         continue;
1132                 id = mdesc_get_property(md, target, "id", NULL);
1133                 if (*id < nr_cpu_ids)
1134                         cpumask_set_cpu(*id, mask);
1135         }
1136 }
1137
1138 static struct mdesc_mlgroup * __init find_mlgroup(u64 node)
1139 {
1140         int i;
1141
1142         for (i = 0; i < num_mlgroups; i++) {
1143                 struct mdesc_mlgroup *m = &mlgroups[i];
1144                 if (m->node == node)
1145                         return m;
1146         }
1147         return NULL;
1148 }
1149
1150 int __node_distance(int from, int to)
1151 {
1152         if ((from >= MAX_NUMNODES) || (to >= MAX_NUMNODES)) {
1153                 pr_warn("Returning default NUMA distance value for %d->%d\n",
1154                         from, to);
1155                 return (from == to) ? LOCAL_DISTANCE : REMOTE_DISTANCE;
1156         }
1157         return numa_latency[from][to];
1158 }
1159
1160 static int find_best_numa_node_for_mlgroup(struct mdesc_mlgroup *grp)
1161 {
1162         int i;
1163
1164         for (i = 0; i < MAX_NUMNODES; i++) {
1165                 struct node_mem_mask *n = &node_masks[i];
1166
1167                 if ((grp->mask == n->mask) && (grp->match == n->val))
1168                         break;
1169         }
1170         return i;
1171 }
1172
1173 static void find_numa_latencies_for_group(struct mdesc_handle *md, u64 grp,
1174                                           int index)
1175 {
1176         u64 arc;
1177
1178         mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1179                 int tnode;
1180                 u64 target = mdesc_arc_target(md, arc);
1181                 struct mdesc_mlgroup *m = find_mlgroup(target);
1182
1183                 if (!m)
1184                         continue;
1185                 tnode = find_best_numa_node_for_mlgroup(m);
1186                 if (tnode == MAX_NUMNODES)
1187                         continue;
1188                 numa_latency[index][tnode] = m->latency;
1189         }
1190 }
1191
1192 static int __init numa_attach_mlgroup(struct mdesc_handle *md, u64 grp,
1193                                       int index)
1194 {
1195         struct mdesc_mlgroup *candidate = NULL;
1196         u64 arc, best_latency = ~(u64)0;
1197         struct node_mem_mask *n;
1198
1199         mdesc_for_each_arc(arc, md, grp, MDESC_ARC_TYPE_FWD) {
1200                 u64 target = mdesc_arc_target(md, arc);
1201                 struct mdesc_mlgroup *m = find_mlgroup(target);
1202                 if (!m)
1203                         continue;
1204                 if (m->latency < best_latency) {
1205                         candidate = m;
1206                         best_latency = m->latency;
1207                 }
1208         }
1209         if (!candidate)
1210                 return -ENOENT;
1211
1212         if (num_node_masks != index) {
1213                 printk(KERN_ERR "Inconsistent NUMA state, "
1214                        "index[%d] != num_node_masks[%d]\n",
1215                        index, num_node_masks);
1216                 return -EINVAL;
1217         }
1218
1219         n = &node_masks[num_node_masks++];
1220
1221         n->mask = candidate->mask;
1222         n->val = candidate->match;
1223
1224         numadbg("NUMA NODE[%d]: mask[%lx] val[%lx] (latency[%llx])\n",
1225                 index, n->mask, n->val, candidate->latency);
1226
1227         return 0;
1228 }
1229
1230 static int __init numa_parse_mdesc_group(struct mdesc_handle *md, u64 grp,
1231                                          int index)
1232 {
1233         cpumask_t mask;
1234         int cpu;
1235
1236         numa_parse_mdesc_group_cpus(md, grp, &mask);
1237
1238         for_each_cpu(cpu, &mask)
1239                 numa_cpu_lookup_table[cpu] = index;
1240         cpumask_copy(&numa_cpumask_lookup_table[index], &mask);
1241
1242         if (numa_debug) {
1243                 printk(KERN_INFO "NUMA GROUP[%d]: cpus [ ", index);
1244                 for_each_cpu(cpu, &mask)
1245                         printk("%d ", cpu);
1246                 printk("]\n");
1247         }
1248
1249         return numa_attach_mlgroup(md, grp, index);
1250 }
1251
1252 static int __init numa_parse_mdesc(void)
1253 {
1254         struct mdesc_handle *md = mdesc_grab();
1255         int i, j, err, count;
1256         u64 node;
1257
1258         node = mdesc_node_by_name(md, MDESC_NODE_NULL, "latency-groups");
1259         if (node == MDESC_NODE_NULL) {
1260                 mdesc_release(md);
1261                 return -ENOENT;
1262         }
1263
1264         err = grab_mblocks(md);
1265         if (err < 0)
1266                 goto out;
1267
1268         err = grab_mlgroups(md);
1269         if (err < 0)
1270                 goto out;
1271
1272         count = 0;
1273         mdesc_for_each_node_by_name(md, node, "group") {
1274                 err = numa_parse_mdesc_group(md, node, count);
1275                 if (err < 0)
1276                         break;
1277                 count++;
1278         }
1279
1280         count = 0;
1281         mdesc_for_each_node_by_name(md, node, "group") {
1282                 find_numa_latencies_for_group(md, node, count);
1283                 count++;
1284         }
1285
1286         /* Normalize numa latency matrix according to ACPI SLIT spec. */
1287         for (i = 0; i < MAX_NUMNODES; i++) {
1288                 u64 self_latency = numa_latency[i][i];
1289
1290                 for (j = 0; j < MAX_NUMNODES; j++) {
1291                         numa_latency[i][j] =
1292                                 (numa_latency[i][j] * LOCAL_DISTANCE) /
1293                                 self_latency;
1294                 }
1295         }
1296
1297         add_node_ranges();
1298
1299         for (i = 0; i < num_node_masks; i++) {
1300                 allocate_node_data(i);
1301                 node_set_online(i);
1302         }
1303
1304         err = 0;
1305 out:
1306         mdesc_release(md);
1307         return err;
1308 }
1309
1310 static int __init numa_parse_jbus(void)
1311 {
1312         unsigned long cpu, index;
1313
1314         /* NUMA node id is encoded in bits 36 and higher, and there is
1315          * a 1-to-1 mapping from CPU ID to NUMA node ID.
1316          */
1317         index = 0;
1318         for_each_present_cpu(cpu) {
1319                 numa_cpu_lookup_table[cpu] = index;
1320                 cpumask_copy(&numa_cpumask_lookup_table[index], cpumask_of(cpu));
1321                 node_masks[index].mask = ~((1UL << 36UL) - 1UL);
1322                 node_masks[index].val = cpu << 36UL;
1323
1324                 index++;
1325         }
1326         num_node_masks = index;
1327
1328         add_node_ranges();
1329
1330         for (index = 0; index < num_node_masks; index++) {
1331                 allocate_node_data(index);
1332                 node_set_online(index);
1333         }
1334
1335         return 0;
1336 }
1337
1338 static int __init numa_parse_sun4u(void)
1339 {
1340         if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1341                 unsigned long ver;
1342
1343                 __asm__ ("rdpr %%ver, %0" : "=r" (ver));
1344                 if ((ver >> 32UL) == __JALAPENO_ID ||
1345                     (ver >> 32UL) == __SERRANO_ID)
1346                         return numa_parse_jbus();
1347         }
1348         return -1;
1349 }
1350
1351 static int __init bootmem_init_numa(void)
1352 {
1353         int i, j;
1354         int err = -1;
1355
1356         numadbg("bootmem_init_numa()\n");
1357
1358         /* Some sane defaults for numa latency values */
1359         for (i = 0; i < MAX_NUMNODES; i++) {
1360                 for (j = 0; j < MAX_NUMNODES; j++)
1361                         numa_latency[i][j] = (i == j) ?
1362                                 LOCAL_DISTANCE : REMOTE_DISTANCE;
1363         }
1364
1365         if (numa_enabled) {
1366                 if (tlb_type == hypervisor)
1367                         err = numa_parse_mdesc();
1368                 else
1369                         err = numa_parse_sun4u();
1370         }
1371         return err;
1372 }
1373
1374 #else
1375
1376 static int bootmem_init_numa(void)
1377 {
1378         return -1;
1379 }
1380
1381 #endif
1382
1383 static void __init bootmem_init_nonnuma(void)
1384 {
1385         unsigned long top_of_ram = memblock_end_of_DRAM();
1386         unsigned long total_ram = memblock_phys_mem_size();
1387
1388         numadbg("bootmem_init_nonnuma()\n");
1389
1390         printk(KERN_INFO "Top of RAM: 0x%lx, Total RAM: 0x%lx\n",
1391                top_of_ram, total_ram);
1392         printk(KERN_INFO "Memory hole size: %ldMB\n",
1393                (top_of_ram - total_ram) >> 20);
1394
1395         init_node_masks_nonnuma();
1396         memblock_set_node(0, (phys_addr_t)ULLONG_MAX, &memblock.memory, 0);
1397         allocate_node_data(0);
1398         node_set_online(0);
1399 }
1400
1401 static unsigned long __init bootmem_init(unsigned long phys_base)
1402 {
1403         unsigned long end_pfn;
1404
1405         end_pfn = memblock_end_of_DRAM() >> PAGE_SHIFT;
1406         max_pfn = max_low_pfn = end_pfn;
1407         min_low_pfn = (phys_base >> PAGE_SHIFT);
1408
1409         if (bootmem_init_numa() < 0)
1410                 bootmem_init_nonnuma();
1411
1412         /* Dump memblock with node info. */
1413         memblock_dump_all();
1414
1415         /* XXX cpu notifier XXX */
1416
1417         sparse_memory_present_with_active_regions(MAX_NUMNODES);
1418         sparse_init();
1419
1420         return end_pfn;
1421 }
1422
1423 static struct linux_prom64_registers pall[MAX_BANKS] __initdata;
1424 static int pall_ents __initdata;
1425
1426 static unsigned long max_phys_bits = 40;
1427
1428 bool kern_addr_valid(unsigned long addr)
1429 {
1430         pgd_t *pgd;
1431         pud_t *pud;
1432         pmd_t *pmd;
1433         pte_t *pte;
1434
1435         if ((long)addr < 0L) {
1436                 unsigned long pa = __pa(addr);
1437
1438                 if ((addr >> max_phys_bits) != 0UL)
1439                         return false;
1440
1441                 return pfn_valid(pa >> PAGE_SHIFT);
1442         }
1443
1444         if (addr >= (unsigned long) KERNBASE &&
1445             addr < (unsigned long)&_end)
1446                 return true;
1447
1448         pgd = pgd_offset_k(addr);
1449         if (pgd_none(*pgd))
1450                 return 0;
1451
1452         pud = pud_offset(pgd, addr);
1453         if (pud_none(*pud))
1454                 return 0;
1455
1456         if (pud_large(*pud))
1457                 return pfn_valid(pud_pfn(*pud));
1458
1459         pmd = pmd_offset(pud, addr);
1460         if (pmd_none(*pmd))
1461                 return 0;
1462
1463         if (pmd_large(*pmd))
1464                 return pfn_valid(pmd_pfn(*pmd));
1465
1466         pte = pte_offset_kernel(pmd, addr);
1467         if (pte_none(*pte))
1468                 return 0;
1469
1470         return pfn_valid(pte_pfn(*pte));
1471 }
1472 EXPORT_SYMBOL(kern_addr_valid);
1473
1474 static unsigned long __ref kernel_map_hugepud(unsigned long vstart,
1475                                               unsigned long vend,
1476                                               pud_t *pud)
1477 {
1478         const unsigned long mask16gb = (1UL << 34) - 1UL;
1479         u64 pte_val = vstart;
1480
1481         /* Each PUD is 8GB */
1482         if ((vstart & mask16gb) ||
1483             (vend - vstart <= mask16gb)) {
1484                 pte_val ^= kern_linear_pte_xor[2];
1485                 pud_val(*pud) = pte_val | _PAGE_PUD_HUGE;
1486
1487                 return vstart + PUD_SIZE;
1488         }
1489
1490         pte_val ^= kern_linear_pte_xor[3];
1491         pte_val |= _PAGE_PUD_HUGE;
1492
1493         vend = vstart + mask16gb + 1UL;
1494         while (vstart < vend) {
1495                 pud_val(*pud) = pte_val;
1496
1497                 pte_val += PUD_SIZE;
1498                 vstart += PUD_SIZE;
1499                 pud++;
1500         }
1501         return vstart;
1502 }
1503
1504 static bool kernel_can_map_hugepud(unsigned long vstart, unsigned long vend,
1505                                    bool guard)
1506 {
1507         if (guard && !(vstart & ~PUD_MASK) && (vend - vstart) >= PUD_SIZE)
1508                 return true;
1509
1510         return false;
1511 }
1512
1513 static unsigned long __ref kernel_map_hugepmd(unsigned long vstart,
1514                                               unsigned long vend,
1515                                               pmd_t *pmd)
1516 {
1517         const unsigned long mask256mb = (1UL << 28) - 1UL;
1518         const unsigned long mask2gb = (1UL << 31) - 1UL;
1519         u64 pte_val = vstart;
1520
1521         /* Each PMD is 8MB */
1522         if ((vstart & mask256mb) ||
1523             (vend - vstart <= mask256mb)) {
1524                 pte_val ^= kern_linear_pte_xor[0];
1525                 pmd_val(*pmd) = pte_val | _PAGE_PMD_HUGE;
1526
1527                 return vstart + PMD_SIZE;
1528         }
1529
1530         if ((vstart & mask2gb) ||
1531             (vend - vstart <= mask2gb)) {
1532                 pte_val ^= kern_linear_pte_xor[1];
1533                 pte_val |= _PAGE_PMD_HUGE;
1534                 vend = vstart + mask256mb + 1UL;
1535         } else {
1536                 pte_val ^= kern_linear_pte_xor[2];
1537                 pte_val |= _PAGE_PMD_HUGE;
1538                 vend = vstart + mask2gb + 1UL;
1539         }
1540
1541         while (vstart < vend) {
1542                 pmd_val(*pmd) = pte_val;
1543
1544                 pte_val += PMD_SIZE;
1545                 vstart += PMD_SIZE;
1546                 pmd++;
1547         }
1548
1549         return vstart;
1550 }
1551
1552 static bool kernel_can_map_hugepmd(unsigned long vstart, unsigned long vend,
1553                                    bool guard)
1554 {
1555         if (guard && !(vstart & ~PMD_MASK) && (vend - vstart) >= PMD_SIZE)
1556                 return true;
1557
1558         return false;
1559 }
1560
1561 static unsigned long __ref kernel_map_range(unsigned long pstart,
1562                                             unsigned long pend, pgprot_t prot,
1563                                             bool use_huge)
1564 {
1565         unsigned long vstart = PAGE_OFFSET + pstart;
1566         unsigned long vend = PAGE_OFFSET + pend;
1567         unsigned long alloc_bytes = 0UL;
1568
1569         if ((vstart & ~PAGE_MASK) || (vend & ~PAGE_MASK)) {
1570                 prom_printf("kernel_map: Unaligned physmem[%lx:%lx]\n",
1571                             vstart, vend);
1572                 prom_halt();
1573         }
1574
1575         while (vstart < vend) {
1576                 unsigned long this_end, paddr = __pa(vstart);
1577                 pgd_t *pgd = pgd_offset_k(vstart);
1578                 pud_t *pud;
1579                 pmd_t *pmd;
1580                 pte_t *pte;
1581
1582                 if (pgd_none(*pgd)) {
1583                         pud_t *new;
1584
1585                         new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1586                         alloc_bytes += PAGE_SIZE;
1587                         pgd_populate(&init_mm, pgd, new);
1588                 }
1589                 pud = pud_offset(pgd, vstart);
1590                 if (pud_none(*pud)) {
1591                         pmd_t *new;
1592
1593                         if (kernel_can_map_hugepud(vstart, vend, use_huge)) {
1594                                 vstart = kernel_map_hugepud(vstart, vend, pud);
1595                                 continue;
1596                         }
1597                         new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1598                         alloc_bytes += PAGE_SIZE;
1599                         pud_populate(&init_mm, pud, new);
1600                 }
1601
1602                 pmd = pmd_offset(pud, vstart);
1603                 if (pmd_none(*pmd)) {
1604                         pte_t *new;
1605
1606                         if (kernel_can_map_hugepmd(vstart, vend, use_huge)) {
1607                                 vstart = kernel_map_hugepmd(vstart, vend, pmd);
1608                                 continue;
1609                         }
1610                         new = __alloc_bootmem(PAGE_SIZE, PAGE_SIZE, PAGE_SIZE);
1611                         alloc_bytes += PAGE_SIZE;
1612                         pmd_populate_kernel(&init_mm, pmd, new);
1613                 }
1614
1615                 pte = pte_offset_kernel(pmd, vstart);
1616                 this_end = (vstart + PMD_SIZE) & PMD_MASK;
1617                 if (this_end > vend)
1618                         this_end = vend;
1619
1620                 while (vstart < this_end) {
1621                         pte_val(*pte) = (paddr | pgprot_val(prot));
1622
1623                         vstart += PAGE_SIZE;
1624                         paddr += PAGE_SIZE;
1625                         pte++;
1626                 }
1627         }
1628
1629         return alloc_bytes;
1630 }
1631
1632 static void __init flush_all_kernel_tsbs(void)
1633 {
1634         int i;
1635
1636         for (i = 0; i < KERNEL_TSB_NENTRIES; i++) {
1637                 struct tsb *ent = &swapper_tsb[i];
1638
1639                 ent->tag = (1UL << TSB_TAG_INVALID_BIT);
1640         }
1641 #ifndef CONFIG_DEBUG_PAGEALLOC
1642         for (i = 0; i < KERNEL_TSB4M_NENTRIES; i++) {
1643                 struct tsb *ent = &swapper_4m_tsb[i];
1644
1645                 ent->tag = (1UL << TSB_TAG_INVALID_BIT);
1646         }
1647 #endif
1648 }
1649
1650 extern unsigned int kvmap_linear_patch[1];
1651
1652 static void __init kernel_physical_mapping_init(void)
1653 {
1654         unsigned long i, mem_alloced = 0UL;
1655         bool use_huge = true;
1656
1657 #ifdef CONFIG_DEBUG_PAGEALLOC
1658         use_huge = false;
1659 #endif
1660         for (i = 0; i < pall_ents; i++) {
1661                 unsigned long phys_start, phys_end;
1662
1663                 phys_start = pall[i].phys_addr;
1664                 phys_end = phys_start + pall[i].reg_size;
1665
1666                 mem_alloced += kernel_map_range(phys_start, phys_end,
1667                                                 PAGE_KERNEL, use_huge);
1668         }
1669
1670         printk("Allocated %ld bytes for kernel page tables.\n",
1671                mem_alloced);
1672
1673         kvmap_linear_patch[0] = 0x01000000; /* nop */
1674         flushi(&kvmap_linear_patch[0]);
1675
1676         flush_all_kernel_tsbs();
1677
1678         __flush_tlb_all();
1679 }
1680
1681 #ifdef CONFIG_DEBUG_PAGEALLOC
1682 void __kernel_map_pages(struct page *page, int numpages, int enable)
1683 {
1684         unsigned long phys_start = page_to_pfn(page) << PAGE_SHIFT;
1685         unsigned long phys_end = phys_start + (numpages * PAGE_SIZE);
1686
1687         kernel_map_range(phys_start, phys_end,
1688                          (enable ? PAGE_KERNEL : __pgprot(0)), false);
1689
1690         flush_tsb_kernel_range(PAGE_OFFSET + phys_start,
1691                                PAGE_OFFSET + phys_end);
1692
1693         /* we should perform an IPI and flush all tlbs,
1694          * but that can deadlock->flush only current cpu.
1695          */
1696         __flush_tlb_kernel_range(PAGE_OFFSET + phys_start,
1697                                  PAGE_OFFSET + phys_end);
1698 }
1699 #endif
1700
1701 unsigned long __init find_ecache_flush_span(unsigned long size)
1702 {
1703         int i;
1704
1705         for (i = 0; i < pavail_ents; i++) {
1706                 if (pavail[i].reg_size >= size)
1707                         return pavail[i].phys_addr;
1708         }
1709
1710         return ~0UL;
1711 }
1712
1713 unsigned long PAGE_OFFSET;
1714 EXPORT_SYMBOL(PAGE_OFFSET);
1715
1716 unsigned long VMALLOC_END   = 0x0000010000000000UL;
1717 EXPORT_SYMBOL(VMALLOC_END);
1718
1719 unsigned long sparc64_va_hole_top =    0xfffff80000000000UL;
1720 unsigned long sparc64_va_hole_bottom = 0x0000080000000000UL;
1721
1722 static void __init setup_page_offset(void)
1723 {
1724         if (tlb_type == cheetah || tlb_type == cheetah_plus) {
1725                 /* Cheetah/Panther support a full 64-bit virtual
1726                  * address, so we can use all that our page tables
1727                  * support.
1728                  */
1729                 sparc64_va_hole_top =    0xfff0000000000000UL;
1730                 sparc64_va_hole_bottom = 0x0010000000000000UL;
1731
1732                 max_phys_bits = 42;
1733         } else if (tlb_type == hypervisor) {
1734                 switch (sun4v_chip_type) {
1735                 case SUN4V_CHIP_NIAGARA1:
1736                 case SUN4V_CHIP_NIAGARA2:
1737                         /* T1 and T2 support 48-bit virtual addresses.  */
1738                         sparc64_va_hole_top =    0xffff800000000000UL;
1739                         sparc64_va_hole_bottom = 0x0000800000000000UL;
1740
1741                         max_phys_bits = 39;
1742                         break;
1743                 case SUN4V_CHIP_NIAGARA3:
1744                         /* T3 supports 48-bit virtual addresses.  */
1745                         sparc64_va_hole_top =    0xffff800000000000UL;
1746                         sparc64_va_hole_bottom = 0x0000800000000000UL;
1747
1748                         max_phys_bits = 43;
1749                         break;
1750                 case SUN4V_CHIP_NIAGARA4:
1751                 case SUN4V_CHIP_NIAGARA5:
1752                 case SUN4V_CHIP_SPARC64X:
1753                 case SUN4V_CHIP_SPARC_M6:
1754                         /* T4 and later support 52-bit virtual addresses.  */
1755                         sparc64_va_hole_top =    0xfff8000000000000UL;
1756                         sparc64_va_hole_bottom = 0x0008000000000000UL;
1757                         max_phys_bits = 47;
1758                         break;
1759                 case SUN4V_CHIP_SPARC_M7:
1760                 case SUN4V_CHIP_SPARC_SN:
1761                 default:
1762                         /* M7 and later support 52-bit virtual addresses.  */
1763                         sparc64_va_hole_top =    0xfff8000000000000UL;
1764                         sparc64_va_hole_bottom = 0x0008000000000000UL;
1765                         max_phys_bits = 49;
1766                         break;
1767                 }
1768         }
1769
1770         if (max_phys_bits > MAX_PHYS_ADDRESS_BITS) {
1771                 prom_printf("MAX_PHYS_ADDRESS_BITS is too small, need %lu\n",
1772                             max_phys_bits);
1773                 prom_halt();
1774         }
1775
1776         PAGE_OFFSET = sparc64_va_hole_top;
1777         VMALLOC_END = ((sparc64_va_hole_bottom >> 1) +
1778                        (sparc64_va_hole_bottom >> 2));
1779
1780         pr_info("MM: PAGE_OFFSET is 0x%016lx (max_phys_bits == %lu)\n",
1781                 PAGE_OFFSET, max_phys_bits);
1782         pr_info("MM: VMALLOC [0x%016lx --> 0x%016lx]\n",
1783                 VMALLOC_START, VMALLOC_END);
1784         pr_info("MM: VMEMMAP [0x%016lx --> 0x%016lx]\n",
1785                 VMEMMAP_BASE, VMEMMAP_BASE << 1);
1786 }
1787
1788 static void __init tsb_phys_patch(void)
1789 {
1790         struct tsb_ldquad_phys_patch_entry *pquad;
1791         struct tsb_phys_patch_entry *p;
1792
1793         pquad = &__tsb_ldquad_phys_patch;
1794         while (pquad < &__tsb_ldquad_phys_patch_end) {
1795                 unsigned long addr = pquad->addr;
1796
1797                 if (tlb_type == hypervisor)
1798                         *(unsigned int *) addr = pquad->sun4v_insn;
1799                 else
1800                         *(unsigned int *) addr = pquad->sun4u_insn;
1801                 wmb();
1802                 __asm__ __volatile__("flush     %0"
1803                                      : /* no outputs */
1804                                      : "r" (addr));
1805
1806                 pquad++;
1807         }
1808
1809         p = &__tsb_phys_patch;
1810         while (p < &__tsb_phys_patch_end) {
1811                 unsigned long addr = p->addr;
1812
1813                 *(unsigned int *) addr = p->insn;
1814                 wmb();
1815                 __asm__ __volatile__("flush     %0"
1816                                      : /* no outputs */
1817                                      : "r" (addr));
1818
1819                 p++;
1820         }
1821 }
1822
1823 /* Don't mark as init, we give this to the Hypervisor.  */
1824 #ifndef CONFIG_DEBUG_PAGEALLOC
1825 #define NUM_KTSB_DESCR  2
1826 #else
1827 #define NUM_KTSB_DESCR  1
1828 #endif
1829 static struct hv_tsb_descr ktsb_descr[NUM_KTSB_DESCR];
1830
1831 /* The swapper TSBs are loaded with a base sequence of:
1832  *
1833  *      sethi   %uhi(SYMBOL), REG1
1834  *      sethi   %hi(SYMBOL), REG2
1835  *      or      REG1, %ulo(SYMBOL), REG1
1836  *      or      REG2, %lo(SYMBOL), REG2
1837  *      sllx    REG1, 32, REG1
1838  *      or      REG1, REG2, REG1
1839  *
1840  * When we use physical addressing for the TSB accesses, we patch the
1841  * first four instructions in the above sequence.
1842  */
1843
1844 static void patch_one_ktsb_phys(unsigned int *start, unsigned int *end, unsigned long pa)
1845 {
1846         unsigned long high_bits, low_bits;
1847
1848         high_bits = (pa >> 32) & 0xffffffff;
1849         low_bits = (pa >> 0) & 0xffffffff;
1850
1851         while (start < end) {
1852                 unsigned int *ia = (unsigned int *)(unsigned long)*start;
1853
1854                 ia[0] = (ia[0] & ~0x3fffff) | (high_bits >> 10);
1855                 __asm__ __volatile__("flush     %0" : : "r" (ia));
1856
1857                 ia[1] = (ia[1] & ~0x3fffff) | (low_bits >> 10);
1858                 __asm__ __volatile__("flush     %0" : : "r" (ia + 1));
1859
1860                 ia[2] = (ia[2] & ~0x1fff) | (high_bits & 0x3ff);
1861                 __asm__ __volatile__("flush     %0" : : "r" (ia + 2));
1862
1863                 ia[3] = (ia[3] & ~0x1fff) | (low_bits & 0x3ff);
1864                 __asm__ __volatile__("flush     %0" : : "r" (ia + 3));
1865
1866                 start++;
1867         }
1868 }
1869
1870 static void ktsb_phys_patch(void)
1871 {
1872         extern unsigned int __swapper_tsb_phys_patch;
1873         extern unsigned int __swapper_tsb_phys_patch_end;
1874         unsigned long ktsb_pa;
1875
1876         ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
1877         patch_one_ktsb_phys(&__swapper_tsb_phys_patch,
1878                             &__swapper_tsb_phys_patch_end, ktsb_pa);
1879 #ifndef CONFIG_DEBUG_PAGEALLOC
1880         {
1881         extern unsigned int __swapper_4m_tsb_phys_patch;
1882         extern unsigned int __swapper_4m_tsb_phys_patch_end;
1883         ktsb_pa = (kern_base +
1884                    ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
1885         patch_one_ktsb_phys(&__swapper_4m_tsb_phys_patch,
1886                             &__swapper_4m_tsb_phys_patch_end, ktsb_pa);
1887         }
1888 #endif
1889 }
1890
1891 static void __init sun4v_ktsb_init(void)
1892 {
1893         unsigned long ktsb_pa;
1894
1895         /* First KTSB for PAGE_SIZE mappings.  */
1896         ktsb_pa = kern_base + ((unsigned long)&swapper_tsb[0] - KERNBASE);
1897
1898         switch (PAGE_SIZE) {
1899         case 8 * 1024:
1900         default:
1901                 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_8K;
1902                 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_8K;
1903                 break;
1904
1905         case 64 * 1024:
1906                 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_64K;
1907                 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_64K;
1908                 break;
1909
1910         case 512 * 1024:
1911                 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_512K;
1912                 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_512K;
1913                 break;
1914
1915         case 4 * 1024 * 1024:
1916                 ktsb_descr[0].pgsz_idx = HV_PGSZ_IDX_4MB;
1917                 ktsb_descr[0].pgsz_mask = HV_PGSZ_MASK_4MB;
1918                 break;
1919         }
1920
1921         ktsb_descr[0].assoc = 1;
1922         ktsb_descr[0].num_ttes = KERNEL_TSB_NENTRIES;
1923         ktsb_descr[0].ctx_idx = 0;
1924         ktsb_descr[0].tsb_base = ktsb_pa;
1925         ktsb_descr[0].resv = 0;
1926
1927 #ifndef CONFIG_DEBUG_PAGEALLOC
1928         /* Second KTSB for 4MB/256MB/2GB/16GB mappings.  */
1929         ktsb_pa = (kern_base +
1930                    ((unsigned long)&swapper_4m_tsb[0] - KERNBASE));
1931
1932         ktsb_descr[1].pgsz_idx = HV_PGSZ_IDX_4MB;
1933         ktsb_descr[1].pgsz_mask = ((HV_PGSZ_MASK_4MB |
1934                                     HV_PGSZ_MASK_256MB |
1935                                     HV_PGSZ_MASK_2GB |
1936                                     HV_PGSZ_MASK_16GB) &
1937                                    cpu_pgsz_mask);
1938         ktsb_descr[1].assoc = 1;
1939         ktsb_descr[1].num_ttes = KERNEL_TSB4M_NENTRIES;
1940         ktsb_descr[1].ctx_idx = 0;
1941         ktsb_descr[1].tsb_base = ktsb_pa;
1942         ktsb_descr[1].resv = 0;
1943 #endif
1944 }
1945
1946 void sun4v_ktsb_register(void)
1947 {
1948         unsigned long pa, ret;
1949
1950         pa = kern_base + ((unsigned long)&ktsb_descr[0] - KERNBASE);
1951
1952         ret = sun4v_mmu_tsb_ctx0(NUM_KTSB_DESCR, pa);
1953         if (ret != 0) {
1954                 prom_printf("hypervisor_mmu_tsb_ctx0[%lx]: "
1955                             "errors with %lx\n", pa, ret);
1956                 prom_halt();
1957         }
1958 }
1959
1960 static void __init sun4u_linear_pte_xor_finalize(void)
1961 {
1962 #ifndef CONFIG_DEBUG_PAGEALLOC
1963         /* This is where we would add Panther support for
1964          * 32MB and 256MB pages.
1965          */
1966 #endif
1967 }
1968
1969 static void __init sun4v_linear_pte_xor_finalize(void)
1970 {
1971         unsigned long pagecv_flag;
1972
1973         /* Bit 9 of TTE is no longer CV bit on M7 processor and it instead
1974          * enables MCD error. Do not set bit 9 on M7 processor.
1975          */
1976         switch (sun4v_chip_type) {
1977         case SUN4V_CHIP_SPARC_M7:
1978         case SUN4V_CHIP_SPARC_SN:
1979                 pagecv_flag = 0x00;
1980                 break;
1981         default:
1982                 pagecv_flag = _PAGE_CV_4V;
1983                 break;
1984         }
1985 #ifndef CONFIG_DEBUG_PAGEALLOC
1986         if (cpu_pgsz_mask & HV_PGSZ_MASK_256MB) {
1987                 kern_linear_pte_xor[1] = (_PAGE_VALID | _PAGE_SZ256MB_4V) ^
1988                         PAGE_OFFSET;
1989                 kern_linear_pte_xor[1] |= (_PAGE_CP_4V | pagecv_flag |
1990                                            _PAGE_P_4V | _PAGE_W_4V);
1991         } else {
1992                 kern_linear_pte_xor[1] = kern_linear_pte_xor[0];
1993         }
1994
1995         if (cpu_pgsz_mask & HV_PGSZ_MASK_2GB) {
1996                 kern_linear_pte_xor[2] = (_PAGE_VALID | _PAGE_SZ2GB_4V) ^
1997                         PAGE_OFFSET;
1998                 kern_linear_pte_xor[2] |= (_PAGE_CP_4V | pagecv_flag |
1999                                            _PAGE_P_4V | _PAGE_W_4V);
2000         } else {
2001                 kern_linear_pte_xor[2] = kern_linear_pte_xor[1];
2002         }
2003
2004         if (cpu_pgsz_mask & HV_PGSZ_MASK_16GB) {
2005                 kern_linear_pte_xor[3] = (_PAGE_VALID | _PAGE_SZ16GB_4V) ^
2006                         PAGE_OFFSET;
2007                 kern_linear_pte_xor[3] |= (_PAGE_CP_4V | pagecv_flag |
2008                                            _PAGE_P_4V | _PAGE_W_4V);
2009         } else {
2010                 kern_linear_pte_xor[3] = kern_linear_pte_xor[2];
2011         }
2012 #endif
2013 }
2014
2015 /* paging_init() sets up the page tables */
2016
2017 static unsigned long last_valid_pfn;
2018
2019 static void sun4u_pgprot_init(void);
2020 static void sun4v_pgprot_init(void);
2021
2022 static phys_addr_t __init available_memory(void)
2023 {
2024         phys_addr_t available = 0ULL;
2025         phys_addr_t pa_start, pa_end;
2026         u64 i;
2027
2028         for_each_free_mem_range(i, NUMA_NO_NODE, MEMBLOCK_NONE, &pa_start,
2029                                 &pa_end, NULL)
2030                 available = available + (pa_end  - pa_start);
2031
2032         return available;
2033 }
2034
2035 #define _PAGE_CACHE_4U  (_PAGE_CP_4U | _PAGE_CV_4U)
2036 #define _PAGE_CACHE_4V  (_PAGE_CP_4V | _PAGE_CV_4V)
2037 #define __DIRTY_BITS_4U  (_PAGE_MODIFIED_4U | _PAGE_WRITE_4U | _PAGE_W_4U)
2038 #define __DIRTY_BITS_4V  (_PAGE_MODIFIED_4V | _PAGE_WRITE_4V | _PAGE_W_4V)
2039 #define __ACCESS_BITS_4U (_PAGE_ACCESSED_4U | _PAGE_READ_4U | _PAGE_R)
2040 #define __ACCESS_BITS_4V (_PAGE_ACCESSED_4V | _PAGE_READ_4V | _PAGE_R)
2041
2042 /* We need to exclude reserved regions. This exclusion will include
2043  * vmlinux and initrd. To be more precise the initrd size could be used to
2044  * compute a new lower limit because it is freed later during initialization.
2045  */
2046 static void __init reduce_memory(phys_addr_t limit_ram)
2047 {
2048         phys_addr_t avail_ram = available_memory();
2049         phys_addr_t pa_start, pa_end;
2050         u64 i;
2051
2052         if (limit_ram >= avail_ram)
2053                 return;
2054
2055         for_each_free_mem_range(i, NUMA_NO_NODE, MEMBLOCK_NONE, &pa_start,
2056                                 &pa_end, NULL) {
2057                 phys_addr_t region_size = pa_end - pa_start;
2058                 phys_addr_t clip_start = pa_start;
2059
2060                 avail_ram = avail_ram - region_size;
2061                 /* Are we consuming too much? */
2062                 if (avail_ram < limit_ram) {
2063                         phys_addr_t give_back = limit_ram - avail_ram;
2064
2065                         region_size = region_size - give_back;
2066                         clip_start = clip_start + give_back;
2067                 }
2068
2069                 memblock_remove(clip_start, region_size);
2070
2071                 if (avail_ram <= limit_ram)
2072                         break;
2073                 i = 0UL;
2074         }
2075 }
2076
2077 void __init paging_init(void)
2078 {
2079         unsigned long end_pfn, shift, phys_base;
2080         unsigned long real_end, i;
2081         int node;
2082
2083         setup_page_offset();
2084
2085         /* These build time checkes make sure that the dcache_dirty_cpu()
2086          * page->flags usage will work.
2087          *
2088          * When a page gets marked as dcache-dirty, we store the
2089          * cpu number starting at bit 32 in the page->flags.  Also,
2090          * functions like clear_dcache_dirty_cpu use the cpu mask
2091          * in 13-bit signed-immediate instruction fields.
2092          */
2093
2094         /*
2095          * Page flags must not reach into upper 32 bits that are used
2096          * for the cpu number
2097          */
2098         BUILD_BUG_ON(NR_PAGEFLAGS > 32);
2099
2100         /*
2101          * The bit fields placed in the high range must not reach below
2102          * the 32 bit boundary. Otherwise we cannot place the cpu field
2103          * at the 32 bit boundary.
2104          */
2105         BUILD_BUG_ON(SECTIONS_WIDTH + NODES_WIDTH + ZONES_WIDTH +
2106                 ilog2(roundup_pow_of_two(NR_CPUS)) > 32);
2107
2108         BUILD_BUG_ON(NR_CPUS > 4096);
2109
2110         kern_base = (prom_boot_mapping_phys_low >> ILOG2_4MB) << ILOG2_4MB;
2111         kern_size = (unsigned long)&_end - (unsigned long)KERNBASE;
2112
2113         /* Invalidate both kernel TSBs.  */
2114         memset(swapper_tsb, 0x40, sizeof(swapper_tsb));
2115 #ifndef CONFIG_DEBUG_PAGEALLOC
2116         memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
2117 #endif
2118
2119         /* TTE.cv bit on sparc v9 occupies the same position as TTE.mcde
2120          * bit on M7 processor. This is a conflicting usage of the same
2121          * bit. Enabling TTE.cv on M7 would turn on Memory Corruption
2122          * Detection error on all pages and this will lead to problems
2123          * later. Kernel does not run with MCD enabled and hence rest
2124          * of the required steps to fully configure memory corruption
2125          * detection are not taken. We need to ensure TTE.mcde is not
2126          * set on M7 processor. Compute the value of cacheability
2127          * flag for use later taking this into consideration.
2128          */
2129         switch (sun4v_chip_type) {
2130         case SUN4V_CHIP_SPARC_M7:
2131         case SUN4V_CHIP_SPARC_SN:
2132                 page_cache4v_flag = _PAGE_CP_4V;
2133                 break;
2134         default:
2135                 page_cache4v_flag = _PAGE_CACHE_4V;
2136                 break;
2137         }
2138
2139         if (tlb_type == hypervisor)
2140                 sun4v_pgprot_init();
2141         else
2142                 sun4u_pgprot_init();
2143
2144         if (tlb_type == cheetah_plus ||
2145             tlb_type == hypervisor) {
2146                 tsb_phys_patch();
2147                 ktsb_phys_patch();
2148         }
2149
2150         if (tlb_type == hypervisor)
2151                 sun4v_patch_tlb_handlers();
2152
2153         /* Find available physical memory...
2154          *
2155          * Read it twice in order to work around a bug in openfirmware.
2156          * The call to grab this table itself can cause openfirmware to
2157          * allocate memory, which in turn can take away some space from
2158          * the list of available memory.  Reading it twice makes sure
2159          * we really do get the final value.
2160          */
2161         read_obp_translations();
2162         read_obp_memory("reg", &pall[0], &pall_ents);
2163         read_obp_memory("available", &pavail[0], &pavail_ents);
2164         read_obp_memory("available", &pavail[0], &pavail_ents);
2165
2166         phys_base = 0xffffffffffffffffUL;
2167         for (i = 0; i < pavail_ents; i++) {
2168                 phys_base = min(phys_base, pavail[i].phys_addr);
2169                 memblock_add(pavail[i].phys_addr, pavail[i].reg_size);
2170         }
2171
2172         memblock_reserve(kern_base, kern_size);
2173
2174         find_ramdisk(phys_base);
2175
2176         if (cmdline_memory_size)
2177                 reduce_memory(cmdline_memory_size);
2178
2179         memblock_allow_resize();
2180         memblock_dump_all();
2181
2182         set_bit(0, mmu_context_bmap);
2183
2184         shift = kern_base + PAGE_OFFSET - ((unsigned long)KERNBASE);
2185
2186         real_end = (unsigned long)_end;
2187         num_kernel_image_mappings = DIV_ROUND_UP(real_end - KERNBASE, 1 << ILOG2_4MB);
2188         printk("Kernel: Using %d locked TLB entries for main kernel image.\n",
2189                num_kernel_image_mappings);
2190
2191         /* Set kernel pgd to upper alias so physical page computations
2192          * work.
2193          */
2194         init_mm.pgd += ((shift) / (sizeof(pgd_t)));
2195
2196         memset(swapper_pg_dir, 0, sizeof(swapper_pg_dir));
2197
2198         inherit_prom_mappings();
2199
2200         /* Ok, we can use our TLB miss and window trap handlers safely.  */
2201         setup_tba();
2202
2203         __flush_tlb_all();
2204
2205         prom_build_devicetree();
2206         of_populate_present_mask();
2207 #ifndef CONFIG_SMP
2208         of_fill_in_cpu_data();
2209 #endif
2210
2211         if (tlb_type == hypervisor) {
2212                 sun4v_mdesc_init();
2213                 mdesc_populate_present_mask(cpu_all_mask);
2214 #ifndef CONFIG_SMP
2215                 mdesc_fill_in_cpu_data(cpu_all_mask);
2216 #endif
2217                 mdesc_get_page_sizes(cpu_all_mask, &cpu_pgsz_mask);
2218
2219                 sun4v_linear_pte_xor_finalize();
2220
2221                 sun4v_ktsb_init();
2222                 sun4v_ktsb_register();
2223         } else {
2224                 unsigned long impl, ver;
2225
2226                 cpu_pgsz_mask = (HV_PGSZ_MASK_8K | HV_PGSZ_MASK_64K |
2227                                  HV_PGSZ_MASK_512K | HV_PGSZ_MASK_4MB);
2228
2229                 __asm__ __volatile__("rdpr %%ver, %0" : "=r" (ver));
2230                 impl = ((ver >> 32) & 0xffff);
2231                 if (impl == PANTHER_IMPL)
2232                         cpu_pgsz_mask |= (HV_PGSZ_MASK_32MB |
2233                                           HV_PGSZ_MASK_256MB);
2234
2235                 sun4u_linear_pte_xor_finalize();
2236         }
2237
2238         /* Flush the TLBs and the 4M TSB so that the updated linear
2239          * pte XOR settings are realized for all mappings.
2240          */
2241         __flush_tlb_all();
2242 #ifndef CONFIG_DEBUG_PAGEALLOC
2243         memset(swapper_4m_tsb, 0x40, sizeof(swapper_4m_tsb));
2244 #endif
2245         __flush_tlb_all();
2246
2247         /* Setup bootmem... */
2248         last_valid_pfn = end_pfn = bootmem_init(phys_base);
2249
2250         /* Once the OF device tree and MDESC have been setup, we know
2251          * the list of possible cpus.  Therefore we can allocate the
2252          * IRQ stacks.
2253          */
2254         for_each_possible_cpu(i) {
2255                 node = cpu_to_node(i);
2256
2257                 softirq_stack[i] = __alloc_bootmem_node(NODE_DATA(node),
2258                                                         THREAD_SIZE,
2259                                                         THREAD_SIZE, 0);
2260                 hardirq_stack[i] = __alloc_bootmem_node(NODE_DATA(node),
2261                                                         THREAD_SIZE,
2262                                                         THREAD_SIZE, 0);
2263         }
2264
2265         kernel_physical_mapping_init();
2266
2267         {
2268                 unsigned long max_zone_pfns[MAX_NR_ZONES];
2269
2270                 memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
2271
2272                 max_zone_pfns[ZONE_NORMAL] = end_pfn;
2273
2274                 free_area_init_nodes(max_zone_pfns);
2275         }
2276
2277         printk("Booting Linux...\n");
2278 }
2279
2280 int page_in_phys_avail(unsigned long paddr)
2281 {
2282         int i;
2283
2284         paddr &= PAGE_MASK;
2285
2286         for (i = 0; i < pavail_ents; i++) {
2287                 unsigned long start, end;
2288
2289                 start = pavail[i].phys_addr;
2290                 end = start + pavail[i].reg_size;
2291
2292                 if (paddr >= start && paddr < end)
2293                         return 1;
2294         }
2295         if (paddr >= kern_base && paddr < (kern_base + kern_size))
2296                 return 1;
2297 #ifdef CONFIG_BLK_DEV_INITRD
2298         if (paddr >= __pa(initrd_start) &&
2299             paddr < __pa(PAGE_ALIGN(initrd_end)))
2300                 return 1;
2301 #endif
2302
2303         return 0;
2304 }
2305
2306 static void __init register_page_bootmem_info(void)
2307 {
2308 #ifdef CONFIG_NEED_MULTIPLE_NODES
2309         int i;
2310
2311         for_each_online_node(i)
2312                 if (NODE_DATA(i)->node_spanned_pages)
2313                         register_page_bootmem_info_node(NODE_DATA(i));
2314 #endif
2315 }
2316 void __init mem_init(void)
2317 {
2318         high_memory = __va(last_valid_pfn << PAGE_SHIFT);
2319
2320         register_page_bootmem_info();
2321         free_all_bootmem();
2322
2323         /*
2324          * Set up the zero page, mark it reserved, so that page count
2325          * is not manipulated when freeing the page from user ptes.
2326          */
2327         mem_map_zero = alloc_pages(GFP_KERNEL|__GFP_ZERO, 0);
2328         if (mem_map_zero == NULL) {
2329                 prom_printf("paging_init: Cannot alloc zero page.\n");
2330                 prom_halt();
2331         }
2332         mark_page_reserved(mem_map_zero);
2333
2334         mem_init_print_info(NULL);
2335
2336         if (tlb_type == cheetah || tlb_type == cheetah_plus)
2337                 cheetah_ecache_flush_init();
2338 }
2339
2340 void free_initmem(void)
2341 {
2342         unsigned long addr, initend;
2343         int do_free = 1;
2344
2345         /* If the physical memory maps were trimmed by kernel command
2346          * line options, don't even try freeing this initmem stuff up.
2347          * The kernel image could have been in the trimmed out region
2348          * and if so the freeing below will free invalid page structs.
2349          */
2350         if (cmdline_memory_size)
2351                 do_free = 0;
2352
2353         /*
2354          * The init section is aligned to 8k in vmlinux.lds. Page align for >8k pagesizes.
2355          */
2356         addr = PAGE_ALIGN((unsigned long)(__init_begin));
2357         initend = (unsigned long)(__init_end) & PAGE_MASK;
2358         for (; addr < initend; addr += PAGE_SIZE) {
2359                 unsigned long page;
2360
2361                 page = (addr +
2362                         ((unsigned long) __va(kern_base)) -
2363                         ((unsigned long) KERNBASE));
2364                 memset((void *)addr, POISON_FREE_INITMEM, PAGE_SIZE);
2365
2366                 if (do_free)
2367                         free_reserved_page(virt_to_page(page));
2368         }
2369 }
2370
2371 #ifdef CONFIG_BLK_DEV_INITRD
2372 void free_initrd_mem(unsigned long start, unsigned long end)
2373 {
2374         free_reserved_area((void *)start, (void *)end, POISON_FREE_INITMEM,
2375                            "initrd");
2376 }
2377 #endif
2378
2379 pgprot_t PAGE_KERNEL __read_mostly;
2380 EXPORT_SYMBOL(PAGE_KERNEL);
2381
2382 pgprot_t PAGE_KERNEL_LOCKED __read_mostly;
2383 pgprot_t PAGE_COPY __read_mostly;
2384
2385 pgprot_t PAGE_SHARED __read_mostly;
2386 EXPORT_SYMBOL(PAGE_SHARED);
2387
2388 unsigned long pg_iobits __read_mostly;
2389
2390 unsigned long _PAGE_IE __read_mostly;
2391 EXPORT_SYMBOL(_PAGE_IE);
2392
2393 unsigned long _PAGE_E __read_mostly;
2394 EXPORT_SYMBOL(_PAGE_E);
2395
2396 unsigned long _PAGE_CACHE __read_mostly;
2397 EXPORT_SYMBOL(_PAGE_CACHE);
2398
2399 #ifdef CONFIG_SPARSEMEM_VMEMMAP
2400 int __meminit vmemmap_populate(unsigned long vstart, unsigned long vend,
2401                                int node)
2402 {
2403         unsigned long pte_base;
2404
2405         pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2406                     _PAGE_CP_4U | _PAGE_CV_4U |
2407                     _PAGE_P_4U | _PAGE_W_4U);
2408         if (tlb_type == hypervisor)
2409                 pte_base = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2410                             page_cache4v_flag | _PAGE_P_4V | _PAGE_W_4V);
2411
2412         pte_base |= _PAGE_PMD_HUGE;
2413
2414         vstart = vstart & PMD_MASK;
2415         vend = ALIGN(vend, PMD_SIZE);
2416         for (; vstart < vend; vstart += PMD_SIZE) {
2417                 pgd_t *pgd = pgd_offset_k(vstart);
2418                 unsigned long pte;
2419                 pud_t *pud;
2420                 pmd_t *pmd;
2421
2422                 if (pgd_none(*pgd)) {
2423                         pud_t *new = vmemmap_alloc_block(PAGE_SIZE, node);
2424
2425                         if (!new)
2426                                 return -ENOMEM;
2427                         pgd_populate(&init_mm, pgd, new);
2428                 }
2429
2430                 pud = pud_offset(pgd, vstart);
2431                 if (pud_none(*pud)) {
2432                         pmd_t *new = vmemmap_alloc_block(PAGE_SIZE, node);
2433
2434                         if (!new)
2435                                 return -ENOMEM;
2436                         pud_populate(&init_mm, pud, new);
2437                 }
2438
2439                 pmd = pmd_offset(pud, vstart);
2440
2441                 pte = pmd_val(*pmd);
2442                 if (!(pte & _PAGE_VALID)) {
2443                         void *block = vmemmap_alloc_block(PMD_SIZE, node);
2444
2445                         if (!block)
2446                                 return -ENOMEM;
2447
2448                         pmd_val(*pmd) = pte_base | __pa(block);
2449                 }
2450         }
2451
2452         return 0;
2453 }
2454
2455 void vmemmap_free(unsigned long start, unsigned long end)
2456 {
2457 }
2458 #endif /* CONFIG_SPARSEMEM_VMEMMAP */
2459
2460 static void prot_init_common(unsigned long page_none,
2461                              unsigned long page_shared,
2462                              unsigned long page_copy,
2463                              unsigned long page_readonly,
2464                              unsigned long page_exec_bit)
2465 {
2466         PAGE_COPY = __pgprot(page_copy);
2467         PAGE_SHARED = __pgprot(page_shared);
2468
2469         protection_map[0x0] = __pgprot(page_none);
2470         protection_map[0x1] = __pgprot(page_readonly & ~page_exec_bit);
2471         protection_map[0x2] = __pgprot(page_copy & ~page_exec_bit);
2472         protection_map[0x3] = __pgprot(page_copy & ~page_exec_bit);
2473         protection_map[0x4] = __pgprot(page_readonly);
2474         protection_map[0x5] = __pgprot(page_readonly);
2475         protection_map[0x6] = __pgprot(page_copy);
2476         protection_map[0x7] = __pgprot(page_copy);
2477         protection_map[0x8] = __pgprot(page_none);
2478         protection_map[0x9] = __pgprot(page_readonly & ~page_exec_bit);
2479         protection_map[0xa] = __pgprot(page_shared & ~page_exec_bit);
2480         protection_map[0xb] = __pgprot(page_shared & ~page_exec_bit);
2481         protection_map[0xc] = __pgprot(page_readonly);
2482         protection_map[0xd] = __pgprot(page_readonly);
2483         protection_map[0xe] = __pgprot(page_shared);
2484         protection_map[0xf] = __pgprot(page_shared);
2485 }
2486
2487 static void __init sun4u_pgprot_init(void)
2488 {
2489         unsigned long page_none, page_shared, page_copy, page_readonly;
2490         unsigned long page_exec_bit;
2491         int i;
2492
2493         PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2494                                 _PAGE_CACHE_4U | _PAGE_P_4U |
2495                                 __ACCESS_BITS_4U | __DIRTY_BITS_4U |
2496                                 _PAGE_EXEC_4U);
2497         PAGE_KERNEL_LOCKED = __pgprot (_PAGE_PRESENT_4U | _PAGE_VALID |
2498                                        _PAGE_CACHE_4U | _PAGE_P_4U |
2499                                        __ACCESS_BITS_4U | __DIRTY_BITS_4U |
2500                                        _PAGE_EXEC_4U | _PAGE_L_4U);
2501
2502         _PAGE_IE = _PAGE_IE_4U;
2503         _PAGE_E = _PAGE_E_4U;
2504         _PAGE_CACHE = _PAGE_CACHE_4U;
2505
2506         pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4U | __DIRTY_BITS_4U |
2507                      __ACCESS_BITS_4U | _PAGE_E_4U);
2508
2509 #ifdef CONFIG_DEBUG_PAGEALLOC
2510         kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET;
2511 #else
2512         kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4U) ^
2513                 PAGE_OFFSET;
2514 #endif
2515         kern_linear_pte_xor[0] |= (_PAGE_CP_4U | _PAGE_CV_4U |
2516                                    _PAGE_P_4U | _PAGE_W_4U);
2517
2518         for (i = 1; i < 4; i++)
2519                 kern_linear_pte_xor[i] = kern_linear_pte_xor[0];
2520
2521         _PAGE_ALL_SZ_BITS =  (_PAGE_SZ4MB_4U | _PAGE_SZ512K_4U |
2522                               _PAGE_SZ64K_4U | _PAGE_SZ8K_4U |
2523                               _PAGE_SZ32MB_4U | _PAGE_SZ256MB_4U);
2524
2525
2526         page_none = _PAGE_PRESENT_4U | _PAGE_ACCESSED_4U | _PAGE_CACHE_4U;
2527         page_shared = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2528                        __ACCESS_BITS_4U | _PAGE_WRITE_4U | _PAGE_EXEC_4U);
2529         page_copy   = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2530                        __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2531         page_readonly   = (_PAGE_VALID | _PAGE_PRESENT_4U | _PAGE_CACHE_4U |
2532                            __ACCESS_BITS_4U | _PAGE_EXEC_4U);
2533
2534         page_exec_bit = _PAGE_EXEC_4U;
2535
2536         prot_init_common(page_none, page_shared, page_copy, page_readonly,
2537                          page_exec_bit);
2538 }
2539
2540 static void __init sun4v_pgprot_init(void)
2541 {
2542         unsigned long page_none, page_shared, page_copy, page_readonly;
2543         unsigned long page_exec_bit;
2544         int i;
2545
2546         PAGE_KERNEL = __pgprot (_PAGE_PRESENT_4V | _PAGE_VALID |
2547                                 page_cache4v_flag | _PAGE_P_4V |
2548                                 __ACCESS_BITS_4V | __DIRTY_BITS_4V |
2549                                 _PAGE_EXEC_4V);
2550         PAGE_KERNEL_LOCKED = PAGE_KERNEL;
2551
2552         _PAGE_IE = _PAGE_IE_4V;
2553         _PAGE_E = _PAGE_E_4V;
2554         _PAGE_CACHE = page_cache4v_flag;
2555
2556 #ifdef CONFIG_DEBUG_PAGEALLOC
2557         kern_linear_pte_xor[0] = _PAGE_VALID ^ PAGE_OFFSET;
2558 #else
2559         kern_linear_pte_xor[0] = (_PAGE_VALID | _PAGE_SZ4MB_4V) ^
2560                 PAGE_OFFSET;
2561 #endif
2562         kern_linear_pte_xor[0] |= (page_cache4v_flag | _PAGE_P_4V |
2563                                    _PAGE_W_4V);
2564
2565         for (i = 1; i < 4; i++)
2566                 kern_linear_pte_xor[i] = kern_linear_pte_xor[0];
2567
2568         pg_iobits = (_PAGE_VALID | _PAGE_PRESENT_4V | __DIRTY_BITS_4V |
2569                      __ACCESS_BITS_4V | _PAGE_E_4V);
2570
2571         _PAGE_ALL_SZ_BITS = (_PAGE_SZ16GB_4V | _PAGE_SZ2GB_4V |
2572                              _PAGE_SZ256MB_4V | _PAGE_SZ32MB_4V |
2573                              _PAGE_SZ4MB_4V | _PAGE_SZ512K_4V |
2574                              _PAGE_SZ64K_4V | _PAGE_SZ8K_4V);
2575
2576         page_none = _PAGE_PRESENT_4V | _PAGE_ACCESSED_4V | page_cache4v_flag;
2577         page_shared = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2578                        __ACCESS_BITS_4V | _PAGE_WRITE_4V | _PAGE_EXEC_4V);
2579         page_copy   = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2580                        __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2581         page_readonly = (_PAGE_VALID | _PAGE_PRESENT_4V | page_cache4v_flag |
2582                          __ACCESS_BITS_4V | _PAGE_EXEC_4V);
2583
2584         page_exec_bit = _PAGE_EXEC_4V;
2585
2586         prot_init_common(page_none, page_shared, page_copy, page_readonly,
2587                          page_exec_bit);
2588 }
2589
2590 unsigned long pte_sz_bits(unsigned long sz)
2591 {
2592         if (tlb_type == hypervisor) {
2593                 switch (sz) {
2594                 case 8 * 1024:
2595                 default:
2596                         return _PAGE_SZ8K_4V;
2597                 case 64 * 1024:
2598                         return _PAGE_SZ64K_4V;
2599                 case 512 * 1024:
2600                         return _PAGE_SZ512K_4V;
2601                 case 4 * 1024 * 1024:
2602                         return _PAGE_SZ4MB_4V;
2603                 }
2604         } else {
2605                 switch (sz) {
2606                 case 8 * 1024:
2607                 default:
2608                         return _PAGE_SZ8K_4U;
2609                 case 64 * 1024:
2610                         return _PAGE_SZ64K_4U;
2611                 case 512 * 1024:
2612                         return _PAGE_SZ512K_4U;
2613                 case 4 * 1024 * 1024:
2614                         return _PAGE_SZ4MB_4U;
2615                 }
2616         }
2617 }
2618
2619 pte_t mk_pte_io(unsigned long page, pgprot_t prot, int space, unsigned long page_size)
2620 {
2621         pte_t pte;
2622
2623         pte_val(pte)  = page | pgprot_val(pgprot_noncached(prot));
2624         pte_val(pte) |= (((unsigned long)space) << 32);
2625         pte_val(pte) |= pte_sz_bits(page_size);
2626
2627         return pte;
2628 }
2629
2630 static unsigned long kern_large_tte(unsigned long paddr)
2631 {
2632         unsigned long val;
2633
2634         val = (_PAGE_VALID | _PAGE_SZ4MB_4U |
2635                _PAGE_CP_4U | _PAGE_CV_4U | _PAGE_P_4U |
2636                _PAGE_EXEC_4U | _PAGE_L_4U | _PAGE_W_4U);
2637         if (tlb_type == hypervisor)
2638                 val = (_PAGE_VALID | _PAGE_SZ4MB_4V |
2639                        page_cache4v_flag | _PAGE_P_4V |
2640                        _PAGE_EXEC_4V | _PAGE_W_4V);
2641
2642         return val | paddr;
2643 }
2644
2645 /* If not locked, zap it. */
2646 void __flush_tlb_all(void)
2647 {
2648         unsigned long pstate;
2649         int i;
2650
2651         __asm__ __volatile__("flushw\n\t"
2652                              "rdpr      %%pstate, %0\n\t"
2653                              "wrpr      %0, %1, %%pstate"
2654                              : "=r" (pstate)
2655                              : "i" (PSTATE_IE));
2656         if (tlb_type == hypervisor) {
2657                 sun4v_mmu_demap_all();
2658         } else if (tlb_type == spitfire) {
2659                 for (i = 0; i < 64; i++) {
2660                         /* Spitfire Errata #32 workaround */
2661                         /* NOTE: Always runs on spitfire, so no
2662                          *       cheetah+ page size encodings.
2663                          */
2664                         __asm__ __volatile__("stxa      %0, [%1] %2\n\t"
2665                                              "flush     %%g6"
2666                                              : /* No outputs */
2667                                              : "r" (0),
2668                                              "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2669
2670                         if (!(spitfire_get_dtlb_data(i) & _PAGE_L_4U)) {
2671                                 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2672                                                      "membar #Sync"
2673                                                      : /* no outputs */
2674                                                      : "r" (TLB_TAG_ACCESS), "i" (ASI_DMMU));
2675                                 spitfire_put_dtlb_data(i, 0x0UL);
2676                         }
2677
2678                         /* Spitfire Errata #32 workaround */
2679                         /* NOTE: Always runs on spitfire, so no
2680                          *       cheetah+ page size encodings.
2681                          */
2682                         __asm__ __volatile__("stxa      %0, [%1] %2\n\t"
2683                                              "flush     %%g6"
2684                                              : /* No outputs */
2685                                              : "r" (0),
2686                                              "r" (PRIMARY_CONTEXT), "i" (ASI_DMMU));
2687
2688                         if (!(spitfire_get_itlb_data(i) & _PAGE_L_4U)) {
2689                                 __asm__ __volatile__("stxa %%g0, [%0] %1\n\t"
2690                                                      "membar #Sync"
2691                                                      : /* no outputs */
2692                                                      : "r" (TLB_TAG_ACCESS), "i" (ASI_IMMU));
2693                                 spitfire_put_itlb_data(i, 0x0UL);
2694                         }
2695                 }
2696         } else if (tlb_type == cheetah || tlb_type == cheetah_plus) {
2697                 cheetah_flush_dtlb_all();
2698                 cheetah_flush_itlb_all();
2699         }
2700         __asm__ __volatile__("wrpr      %0, 0, %%pstate"
2701                              : : "r" (pstate));
2702 }
2703
2704 pte_t *pte_alloc_one_kernel(struct mm_struct *mm,
2705                             unsigned long address)
2706 {
2707         struct page *page = alloc_page(GFP_KERNEL | __GFP_NOTRACK | __GFP_ZERO);
2708         pte_t *pte = NULL;
2709
2710         if (page)
2711                 pte = (pte_t *) page_address(page);
2712
2713         return pte;
2714 }
2715
2716 pgtable_t pte_alloc_one(struct mm_struct *mm,
2717                         unsigned long address)
2718 {
2719         struct page *page = alloc_page(GFP_KERNEL | __GFP_NOTRACK | __GFP_ZERO);
2720         if (!page)
2721                 return NULL;
2722         if (!pgtable_page_ctor(page)) {
2723                 free_hot_cold_page(page, 0);
2724                 return NULL;
2725         }
2726         return (pte_t *) page_address(page);
2727 }
2728
2729 void pte_free_kernel(struct mm_struct *mm, pte_t *pte)
2730 {
2731         free_page((unsigned long)pte);
2732 }
2733
2734 static void __pte_free(pgtable_t pte)
2735 {
2736         struct page *page = virt_to_page(pte);
2737
2738         pgtable_page_dtor(page);
2739         __free_page(page);
2740 }
2741
2742 void pte_free(struct mm_struct *mm, pgtable_t pte)
2743 {
2744         __pte_free(pte);
2745 }
2746
2747 void pgtable_free(void *table, bool is_page)
2748 {
2749         if (is_page)
2750                 __pte_free(table);
2751         else
2752                 kmem_cache_free(pgtable_cache, table);
2753 }
2754
2755 #ifdef CONFIG_TRANSPARENT_HUGEPAGE
2756 void update_mmu_cache_pmd(struct vm_area_struct *vma, unsigned long addr,
2757                           pmd_t *pmd)
2758 {
2759         unsigned long pte, flags;
2760         struct mm_struct *mm;
2761         pmd_t entry = *pmd;
2762
2763         if (!pmd_large(entry) || !pmd_young(entry))
2764                 return;
2765
2766         pte = pmd_val(entry);
2767
2768         /* Don't insert a non-valid PMD into the TSB, we'll deadlock.  */
2769         if (!(pte & _PAGE_VALID))
2770                 return;
2771
2772         /* We are fabricating 8MB pages using 4MB real hw pages.  */
2773         pte |= (addr & (1UL << REAL_HPAGE_SHIFT));
2774
2775         mm = vma->vm_mm;
2776
2777         spin_lock_irqsave(&mm->context.lock, flags);
2778
2779         if (mm->context.tsb_block[MM_TSB_HUGE].tsb != NULL)
2780                 __update_mmu_tsb_insert(mm, MM_TSB_HUGE, REAL_HPAGE_SHIFT,
2781                                         addr, pte);
2782
2783         spin_unlock_irqrestore(&mm->context.lock, flags);
2784 }
2785 #endif /* CONFIG_TRANSPARENT_HUGEPAGE */
2786
2787 #if defined(CONFIG_HUGETLB_PAGE) || defined(CONFIG_TRANSPARENT_HUGEPAGE)
2788 static void context_reload(void *__data)
2789 {
2790         struct mm_struct *mm = __data;
2791
2792         if (mm == current->mm)
2793                 load_secondary_context(mm);
2794 }
2795
2796 void hugetlb_setup(struct pt_regs *regs)
2797 {
2798         struct mm_struct *mm = current->mm;
2799         struct tsb_config *tp;
2800
2801         if (faulthandler_disabled() || !mm) {
2802                 const struct exception_table_entry *entry;
2803
2804                 entry = search_exception_tables(regs->tpc);
2805                 if (entry) {
2806                         regs->tpc = entry->fixup;
2807                         regs->tnpc = regs->tpc + 4;
2808                         return;
2809                 }
2810                 pr_alert("Unexpected HugeTLB setup in atomic context.\n");
2811                 die_if_kernel("HugeTSB in atomic", regs);
2812         }
2813
2814         tp = &mm->context.tsb_block[MM_TSB_HUGE];
2815         if (likely(tp->tsb == NULL))
2816                 tsb_grow(mm, MM_TSB_HUGE, 0);
2817
2818         tsb_context_switch(mm);
2819         smp_tsb_sync(mm);
2820
2821         /* On UltraSPARC-III+ and later, configure the second half of
2822          * the Data-TLB for huge pages.
2823          */
2824         if (tlb_type == cheetah_plus) {
2825                 bool need_context_reload = false;
2826                 unsigned long ctx;
2827
2828                 spin_lock_irq(&ctx_alloc_lock);
2829                 ctx = mm->context.sparc64_ctx_val;
2830                 ctx &= ~CTX_PGSZ_MASK;
2831                 ctx |= CTX_PGSZ_BASE << CTX_PGSZ0_SHIFT;
2832                 ctx |= CTX_PGSZ_HUGE << CTX_PGSZ1_SHIFT;
2833
2834                 if (ctx != mm->context.sparc64_ctx_val) {
2835                         /* When changing the page size fields, we
2836                          * must perform a context flush so that no
2837                          * stale entries match.  This flush must
2838                          * occur with the original context register
2839                          * settings.
2840                          */
2841                         do_flush_tlb_mm(mm);
2842
2843                         /* Reload the context register of all processors
2844                          * also executing in this address space.
2845                          */
2846                         mm->context.sparc64_ctx_val = ctx;
2847                         need_context_reload = true;
2848                 }
2849                 spin_unlock_irq(&ctx_alloc_lock);
2850
2851                 if (need_context_reload)
2852                         on_each_cpu(context_reload, mm, 0);
2853         }
2854 }
2855 #endif
2856
2857 static struct resource code_resource = {
2858         .name   = "Kernel code",
2859         .flags  = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
2860 };
2861
2862 static struct resource data_resource = {
2863         .name   = "Kernel data",
2864         .flags  = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
2865 };
2866
2867 static struct resource bss_resource = {
2868         .name   = "Kernel bss",
2869         .flags  = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM
2870 };
2871
2872 static inline resource_size_t compute_kern_paddr(void *addr)
2873 {
2874         return (resource_size_t) (addr - KERNBASE + kern_base);
2875 }
2876
2877 static void __init kernel_lds_init(void)
2878 {
2879         code_resource.start = compute_kern_paddr(_text);
2880         code_resource.end   = compute_kern_paddr(_etext - 1);
2881         data_resource.start = compute_kern_paddr(_etext);
2882         data_resource.end   = compute_kern_paddr(_edata - 1);
2883         bss_resource.start  = compute_kern_paddr(__bss_start);
2884         bss_resource.end    = compute_kern_paddr(_end - 1);
2885 }
2886
2887 static int __init report_memory(void)
2888 {
2889         int i;
2890         struct resource *res;
2891
2892         kernel_lds_init();
2893
2894         for (i = 0; i < pavail_ents; i++) {
2895                 res = kzalloc(sizeof(struct resource), GFP_KERNEL);
2896
2897                 if (!res) {
2898                         pr_warn("Failed to allocate source.\n");
2899                         break;
2900                 }
2901
2902                 res->name = "System RAM";
2903                 res->start = pavail[i].phys_addr;
2904                 res->end = pavail[i].phys_addr + pavail[i].reg_size - 1;
2905                 res->flags = IORESOURCE_BUSY | IORESOURCE_SYSTEM_RAM;
2906
2907                 if (insert_resource(&iomem_resource, res) < 0) {
2908                         pr_warn("Resource insertion failed.\n");
2909                         break;
2910                 }
2911
2912                 insert_resource(res, &code_resource);
2913                 insert_resource(res, &data_resource);
2914                 insert_resource(res, &bss_resource);
2915         }
2916
2917         return 0;
2918 }
2919 arch_initcall(report_memory);
2920
2921 #ifdef CONFIG_SMP
2922 #define do_flush_tlb_kernel_range       smp_flush_tlb_kernel_range
2923 #else
2924 #define do_flush_tlb_kernel_range       __flush_tlb_kernel_range
2925 #endif
2926
2927 void flush_tlb_kernel_range(unsigned long start, unsigned long end)
2928 {
2929         if (start < HI_OBP_ADDRESS && end > LOW_OBP_ADDRESS) {
2930                 if (start < LOW_OBP_ADDRESS) {
2931                         flush_tsb_kernel_range(start, LOW_OBP_ADDRESS);
2932                         do_flush_tlb_kernel_range(start, LOW_OBP_ADDRESS);
2933                 }
2934                 if (end > HI_OBP_ADDRESS) {
2935                         flush_tsb_kernel_range(HI_OBP_ADDRESS, end);
2936                         do_flush_tlb_kernel_range(HI_OBP_ADDRESS, end);
2937                 }
2938         } else {
2939                 flush_tsb_kernel_range(start, end);
2940                 do_flush_tlb_kernel_range(start, end);
2941         }
2942 }