arch/x86/mm/init_64.c

   1 /*
   2  *  linux/arch/x86_64/mm/init.c
   3  *
   4  *  Copyright (C) 1995  Linus Torvalds
   5  *  Copyright (C) 2000  Pavel Machek <pavel@ucw.cz>
   6  *  Copyright (C) 2002,2003 Andi Kleen <ak@suse.de>
   7  */
   8
   9 #include <linux/signal.h>
  10 #include <linux/sched.h>
  11 #include <linux/kernel.h>
  12 #include <linux/errno.h>
  13 #include <linux/string.h>
  14 #include <linux/types.h>
  15 #include <linux/ptrace.h>
  16 #include <linux/mman.h>
  17 #include <linux/mm.h>
  18 #include <linux/swap.h>
  19 #include <linux/smp.h>
  20 #include <linux/init.h>
  21 #include <linux/initrd.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/bootmem.h>
  24 #include <linux/memblock.h>
  25 #include <linux/proc_fs.h>
  26 #include <linux/pci.h>
  27 #include <linux/pfn.h>
  28 #include <linux/poison.h>
  29 #include <linux/dma-mapping.h>
  30 #include <linux/module.h>
  31 #include <linux/memory.h>
  32 #include <linux/memory_hotplug.h>
  33 #include <linux/nmi.h>
  34 #include <linux/gfp.h>
  35
  36 #include <asm/processor.h>
  37 #include <asm/bios_ebda.h>
  38 #include <asm/uaccess.h>
  39 #include <asm/pgtable.h>
  40 #include <asm/pgalloc.h>
  41 #include <asm/dma.h>
  42 #include <asm/fixmap.h>
  43 #include <asm/e820.h>
  44 #include <asm/apic.h>
  45 #include <asm/tlb.h>
  46 #include <asm/mmu_context.h>
  47 #include <asm/proto.h>
  48 #include <asm/smp.h>
  49 #include <asm/sections.h>
  50 #include <asm/kdebug.h>
  51 #include <asm/numa.h>
  52 #include <asm/cacheflush.h>
  53 #include <asm/init.h>
  54 #include <asm/uv/uv.h>
  55 #include <asm/setup.h>
  56
  57 #include "mm_internal.h"
  58
  59 static int __init parse_direct_gbpages_off(char *arg)
  60 {
  61         direct_gbpages = 0;
  62         return 0;
  63 }
  64 early_param("nogbpages", parse_direct_gbpages_off);
  65
  66 static int __init parse_direct_gbpages_on(char *arg)
  67 {
  68         direct_gbpages = 1;
  69         return 0;
  70 }
  71 early_param("gbpages", parse_direct_gbpages_on);
  72
  73 /*
  74  * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
  75  * physical space so we can cache the place of the first one and move
  76  * around without checking the pgd every time.
  77  */
  78
  79 pteval_t __supported_pte_mask __read_mostly = ~_PAGE_IOMAP;
  80 EXPORT_SYMBOL_GPL(__supported_pte_mask);
  81
  82 int force_personality32;
  83
  84 /*
  85  * noexec32=on|off
  86  * Control non executable heap for 32bit processes.
  87  * To control the stack too use noexec=off
  88  *
  89  * on   PROT_READ does not imply PROT_EXEC for 32-bit processes (default)
  90  * off  PROT_READ implies PROT_EXEC
  91  */
  92 static int __init nonx32_setup(char *str)
  93 {
  94         if (!strcmp(str, "on"))
  95                 force_personality32 &= ~READ_IMPLIES_EXEC;
  96         else if (!strcmp(str, "off"))
  97                 force_personality32 |= READ_IMPLIES_EXEC;
  98         return 1;
  99 }
 100 __setup("noexec32=", nonx32_setup);
 101
 102 /*
 103  * When memory was added/removed make sure all the processes MM have
 104  * suitable PGD entries in the local PGD level page.
 105  */
 106 void sync_global_pgds(unsigned long start, unsigned long end)
 107 {
 108         unsigned long address;
 109
 110         for (address = start; address <= end; address += PGDIR_SIZE) {
 111                 const pgd_t *pgd_ref = pgd_offset_k(address);
 112                 struct page *page;
 113
 114                 if (pgd_none(*pgd_ref))
 115                         continue;
 116
 117                 spin_lock(&pgd_lock);
 118                 list_for_each_entry(page, &pgd_list, lru) {
 119                         pgd_t *pgd;
 120                         spinlock_t *pgt_lock;
 121
 122                         pgd = (pgd_t *)page_address(page) + pgd_index(address);
 123                         /* the pgt_lock only for Xen */
 124                         pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
 125                         spin_lock(pgt_lock);
 126
 127                         if (pgd_none(*pgd))
 128                                 set_pgd(pgd, *pgd_ref);
 129                         else
 130                                 BUG_ON(pgd_page_vaddr(*pgd)
 131                                        != pgd_page_vaddr(*pgd_ref));
 132
 133                         spin_unlock(pgt_lock);
 134                 }
 135                 spin_unlock(&pgd_lock);
 136         }
 137 }
 138
 139 /*
 140  * NOTE: This function is marked __ref because it calls __init function
 141  * (alloc_bootmem_pages). It's safe to do it ONLY when after_bootmem == 0.
 142  */
 143 static __ref void *spp_getpage(void)
 144 {
 145         void *ptr;
 146
 147         if (after_bootmem)
 148                 ptr = (void *) get_zeroed_page(GFP_ATOMIC | __GFP_NOTRACK);
 149         else
 150                 ptr = alloc_bootmem_pages(PAGE_SIZE);
 151
 152         if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
 153                 panic("set_pte_phys: cannot allocate page data %s\n",
 154                         after_bootmem ? "after bootmem" : "");
 155         }
 156
 157         pr_debug("spp_getpage %p\n", ptr);
 158
 159         return ptr;
 160 }
 161
 162 static pud_t *fill_pud(pgd_t *pgd, unsigned long vaddr)
 163 {
 164         if (pgd_none(*pgd)) {
 165                 pud_t *pud = (pud_t *)spp_getpage();
 166                 pgd_populate(&init_mm, pgd, pud);
 167                 if (pud != pud_offset(pgd, 0))
 168                         printk(KERN_ERR "PAGETABLE BUG #00! %p <-> %p\n",
 169                                pud, pud_offset(pgd, 0));
 170         }
 171         return pud_offset(pgd, vaddr);
 172 }
 173
 174 static pmd_t *fill_pmd(pud_t *pud, unsigned long vaddr)
 175 {
 176         if (pud_none(*pud)) {
 177                 pmd_t *pmd = (pmd_t *) spp_getpage();
 178                 pud_populate(&init_mm, pud, pmd);
 179                 if (pmd != pmd_offset(pud, 0))
 180                         printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
 181                                pmd, pmd_offset(pud, 0));
 182         }
 183         return pmd_offset(pud, vaddr);
 184 }
 185
 186 static pte_t *fill_pte(pmd_t *pmd, unsigned long vaddr)
 187 {
 188         if (pmd_none(*pmd)) {
 189                 pte_t *pte = (pte_t *) spp_getpage();
 190                 pmd_populate_kernel(&init_mm, pmd, pte);
 191                 if (pte != pte_offset_kernel(pmd, 0))
 192                         printk(KERN_ERR "PAGETABLE BUG #02!\n");
 193         }
 194         return pte_offset_kernel(pmd, vaddr);
 195 }
 196
 197 void set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
 198 {
 199         pud_t *pud;
 200         pmd_t *pmd;
 201         pte_t *pte;
 202
 203         pud = pud_page + pud_index(vaddr);
 204         pmd = fill_pmd(pud, vaddr);
 205         pte = fill_pte(pmd, vaddr);
 206
 207         set_pte(pte, new_pte);
 208
 209         /*
 210          * It's enough to flush this one mapping.
 211          * (PGE mappings get flushed as well)
 212          */
 213         __flush_tlb_one(vaddr);
 214 }
 215
 216 void set_pte_vaddr(unsigned long vaddr, pte_t pteval)
 217 {
 218         pgd_t *pgd;
 219         pud_t *pud_page;
 220
 221         pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
 222
 223         pgd = pgd_offset_k(vaddr);
 224         if (pgd_none(*pgd)) {
 225                 printk(KERN_ERR
 226                         "PGD FIXMAP MISSING, it should be setup in head.S!\n");
 227                 return;
 228         }
 229         pud_page = (pud_t*)pgd_page_vaddr(*pgd);
 230         set_pte_vaddr_pud(pud_page, vaddr, pteval);
 231 }
 232
 233 pmd_t * __init populate_extra_pmd(unsigned long vaddr)
 234 {
 235         pgd_t *pgd;
 236         pud_t *pud;
 237
 238         pgd = pgd_offset_k(vaddr);
 239         pud = fill_pud(pgd, vaddr);
 240         return fill_pmd(pud, vaddr);
 241 }
 242
 243 pte_t * __init populate_extra_pte(unsigned long vaddr)
 244 {
 245         pmd_t *pmd;
 246
 247         pmd = populate_extra_pmd(vaddr);
 248         return fill_pte(pmd, vaddr);
 249 }
 250
 251 /*
 252  * Create large page table mappings for a range of physical addresses.
 253  */
 254 static void __init __init_extra_mapping(unsigned long phys, unsigned long size,
 255                                                 pgprot_t prot)
 256 {
 257         pgd_t *pgd;
 258         pud_t *pud;
 259         pmd_t *pmd;
 260
 261         BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK));
 262         for (; size; phys += PMD_SIZE, size -= PMD_SIZE) {
 263                 pgd = pgd_offset_k((unsigned long)__va(phys));
 264                 if (pgd_none(*pgd)) {
 265                         pud = (pud_t *) spp_getpage();
 266                         set_pgd(pgd, __pgd(__pa(pud) | _KERNPG_TABLE |
 267                                                 _PAGE_USER));
 268                 }
 269                 pud = pud_offset(pgd, (unsigned long)__va(phys));
 270                 if (pud_none(*pud)) {
 271                         pmd = (pmd_t *) spp_getpage();
 272                         set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE |
 273                                                 _PAGE_USER));
 274                 }
 275                 pmd = pmd_offset(pud, phys);
 276                 BUG_ON(!pmd_none(*pmd));
 277                 set_pmd(pmd, __pmd(phys | pgprot_val(prot)));
 278         }
 279 }
 280
 281 void __init init_extra_mapping_wb(unsigned long phys, unsigned long size)
 282 {
 283         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE);
 284 }
 285
 286 void __init init_extra_mapping_uc(unsigned long phys, unsigned long size)
 287 {
 288         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE_NOCACHE);
 289 }
 290
 291 /*
 292  * The head.S code sets up the kernel high mapping:
 293  *
 294  *   from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
 295  *
 296  * phys_addr holds the negative offset to the kernel, which is added
 297  * to the compile time generated pmds. This results in invalid pmds up
 298  * to the point where we hit the physaddr 0 mapping.
 299  *
 300  * We limit the mappings to the region from _text to _brk_end.  _brk_end
 301  * is rounded up to the 2MB boundary. This catches the invalid pmds as
 302  * well, as they are located before _text:
 303  */
 304 void __init cleanup_highmap(void)
 305 {
 306         unsigned long vaddr = __START_KERNEL_map;
 307         unsigned long vaddr_end = __START_KERNEL_map + (max_pfn_mapped << PAGE_SHIFT);
 308         unsigned long end = roundup((unsigned long)_brk_end, PMD_SIZE) - 1;
 309         pmd_t *pmd = level2_kernel_pgt;
 310
 311         for (; vaddr + PMD_SIZE - 1 < vaddr_end; pmd++, vaddr += PMD_SIZE) {
 312                 if (pmd_none(*pmd))
 313                         continue;
 314                 if (vaddr < (unsigned long) _text || vaddr > end)
 315                         set_pmd(pmd, __pmd(0));
 316         }
 317 }
 318
 319 static unsigned long __meminit
 320 phys_pte_init(pte_t *pte_page, unsigned long addr, unsigned long end,
 321               pgprot_t prot)
 322 {
 323         unsigned long pages = 0, next;
 324         unsigned long last_map_addr = end;
 325         int i;
 326
 327         pte_t *pte = pte_page + pte_index(addr);
 328
 329         for (i = pte_index(addr); i < PTRS_PER_PTE; i++, addr = next, pte++) {
 330                 next = (addr & PAGE_MASK) + PAGE_SIZE;
 331                 if (addr >= end) {
 332                         if (!after_bootmem &&
 333                             !e820_any_mapped(addr & PAGE_MASK, next, E820_RAM) &&
 334                             !e820_any_mapped(addr & PAGE_MASK, next, E820_RESERVED_KERN))
 335                                 set_pte(pte, __pte(0));
 336                         continue;
 337                 }
 338
 339                 /*
 340                  * We will re-use the existing mapping.
 341                  * Xen for example has some special requirements, like mapping
 342                  * pagetable pages as RO. So assume someone who pre-setup
 343                  * these mappings are more intelligent.
 344                  */
 345                 if (pte_val(*pte)) {
 346                         if (!after_bootmem)
 347                                 pages++;
 348                         continue;
 349                 }
 350
 351                 if (0)
 352                         printk("   pte=%p addr=%lx pte=%016lx\n",
 353                                pte, addr, pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL).pte);
 354                 pages++;
 355                 set_pte(pte, pfn_pte(addr >> PAGE_SHIFT, prot));
 356                 last_map_addr = (addr & PAGE_MASK) + PAGE_SIZE;
 357         }
 358
 359         update_page_count(PG_LEVEL_4K, pages);
 360
 361         return last_map_addr;
 362 }
 363
 364 static unsigned long __meminit
 365 phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end,
 366               unsigned long page_size_mask, pgprot_t prot)
 367 {
 368         unsigned long pages = 0, next;
 369         unsigned long last_map_addr = end;
 370
 371         int i = pmd_index(address);
 372
 373         for (; i < PTRS_PER_PMD; i++, address = next) {
 374                 pmd_t *pmd = pmd_page + pmd_index(address);
 375                 pte_t *pte;
 376                 pgprot_t new_prot = prot;
 377
 378                 next = (address & PMD_MASK) + PMD_SIZE;
 379                 if (address >= end) {
 380                         if (!after_bootmem &&
 381                             !e820_any_mapped(address & PMD_MASK, next, E820_RAM) &&
 382                             !e820_any_mapped(address & PMD_MASK, next, E820_RESERVED_KERN))
 383                                 set_pmd(pmd, __pmd(0));
 384                         continue;
 385                 }
 386
 387                 if (pmd_val(*pmd)) {
 388                         if (!pmd_large(*pmd)) {
 389                                 spin_lock(&init_mm.page_table_lock);
 390                                 pte = (pte_t *)pmd_page_vaddr(*pmd);
 391                                 last_map_addr = phys_pte_init(pte, address,
 392                                                                 end, prot);
 393                                 spin_unlock(&init_mm.page_table_lock);
 394                                 continue;
 395                         }
 396                         /*
 397                          * If we are ok with PG_LEVEL_2M mapping, then we will
 398                          * use the existing mapping,
 399                          *
 400                          * Otherwise, we will split the large page mapping but
 401                          * use the same existing protection bits except for
 402                          * large page, so that we don't violate Intel's TLB
 403                          * Application note (317080) which says, while changing
 404                          * the page sizes, new and old translations should
 405                          * not differ with respect to page frame and
 406                          * attributes.
 407                          */
 408                         if (page_size_mask & (1 << PG_LEVEL_2M)) {
 409                                 if (!after_bootmem)
 410                                         pages++;
 411                                 last_map_addr = next;
 412                                 continue;
 413                         }
 414                         new_prot = pte_pgprot(pte_clrhuge(*(pte_t *)pmd));
 415                 }
 416
 417                 if (page_size_mask & (1<<PG_LEVEL_2M)) {
 418                         pages++;
 419                         spin_lock(&init_mm.page_table_lock);
 420                         set_pte((pte_t *)pmd,
 421                                 pfn_pte((address & PMD_MASK) >> PAGE_SHIFT,
 422                                         __pgprot(pgprot_val(prot) | _PAGE_PSE)));
 423                         spin_unlock(&init_mm.page_table_lock);
 424                         last_map_addr = next;
 425                         continue;
 426                 }
 427
 428                 pte = alloc_low_page();
 429                 last_map_addr = phys_pte_init(pte, address, end, new_prot);
 430
 431                 spin_lock(&init_mm.page_table_lock);
 432                 pmd_populate_kernel(&init_mm, pmd, pte);
 433                 spin_unlock(&init_mm.page_table_lock);
 434         }
 435         update_page_count(PG_LEVEL_2M, pages);
 436         return last_map_addr;
 437 }
 438
 439 static unsigned long __meminit
 440 phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end,
 441                          unsigned long page_size_mask)
 442 {
 443         unsigned long pages = 0, next;
 444         unsigned long last_map_addr = end;
 445         int i = pud_index(addr);
 446
 447         for (; i < PTRS_PER_PUD; i++, addr = next) {
 448                 pud_t *pud = pud_page + pud_index(addr);
 449                 pmd_t *pmd;
 450                 pgprot_t prot = PAGE_KERNEL;
 451
 452                 next = (addr & PUD_MASK) + PUD_SIZE;
 453                 if (addr >= end) {
 454                         if (!after_bootmem &&
 455                             !e820_any_mapped(addr & PUD_MASK, next, E820_RAM) &&
 456                             !e820_any_mapped(addr & PUD_MASK, next, E820_RESERVED_KERN))
 457                                 set_pud(pud, __pud(0));
 458                         continue;
 459                 }
 460
 461                 if (pud_val(*pud)) {
 462                         if (!pud_large(*pud)) {
 463                                 pmd = pmd_offset(pud, 0);
 464                                 last_map_addr = phys_pmd_init(pmd, addr, end,
 465                                                          page_size_mask, prot);
 466                                 __flush_tlb_all();
 467                                 continue;
 468                         }
 469                         /*
 470                          * If we are ok with PG_LEVEL_1G mapping, then we will
 471                          * use the existing mapping.
 472                          *
 473                          * Otherwise, we will split the gbpage mapping but use
 474                          * the same existing protection  bits except for large
 475                          * page, so that we don't violate Intel's TLB
 476                          * Application note (317080) which says, while changing
 477                          * the page sizes, new and old translations should
 478                          * not differ with respect to page frame and
 479                          * attributes.
 480                          */
 481                         if (page_size_mask & (1 << PG_LEVEL_1G)) {
 482                                 if (!after_bootmem)
 483                                         pages++;
 484                                 last_map_addr = next;
 485                                 continue;
 486                         }
 487                         prot = pte_pgprot(pte_clrhuge(*(pte_t *)pud));
 488                 }
 489
 490                 if (page_size_mask & (1<<PG_LEVEL_1G)) {
 491                         pages++;
 492                         spin_lock(&init_mm.page_table_lock);
 493                         set_pte((pte_t *)pud,
 494                                 pfn_pte((addr & PUD_MASK) >> PAGE_SHIFT,
 495                                         PAGE_KERNEL_LARGE));
 496                         spin_unlock(&init_mm.page_table_lock);
 497                         last_map_addr = next;
 498                         continue;
 499                 }
 500
 501                 pmd = alloc_low_page();
 502                 last_map_addr = phys_pmd_init(pmd, addr, end, page_size_mask,
 503                                               prot);
 504
 505                 spin_lock(&init_mm.page_table_lock);
 506                 pud_populate(&init_mm, pud, pmd);
 507                 spin_unlock(&init_mm.page_table_lock);
 508         }
 509         __flush_tlb_all();
 510
 511         update_page_count(PG_LEVEL_1G, pages);
 512
 513         return last_map_addr;
 514 }
 515
 516 unsigned long __meminit
 517 kernel_physical_mapping_init(unsigned long start,
 518                              unsigned long end,
 519                              unsigned long page_size_mask)
 520 {
 521         bool pgd_changed = false;
 522         unsigned long next, last_map_addr = end;
 523         unsigned long addr;
 524
 525         start = (unsigned long)__va(start);
 526         end = (unsigned long)__va(end);
 527         addr = start;
 528
 529         for (; start < end; start = next) {
 530                 pgd_t *pgd = pgd_offset_k(start);
 531                 pud_t *pud;
 532
 533                 next = (start + PGDIR_SIZE) & PGDIR_MASK;
 534                 if (next > end)
 535                         next = end;
 536
 537                 if (pgd_val(*pgd)) {
 538                         pud = (pud_t *)pgd_page_vaddr(*pgd);
 539                         last_map_addr = phys_pud_init(pud, __pa(start),
 540                                                  __pa(end), page_size_mask);
 541                         continue;
 542                 }
 543
 544                 pud = alloc_low_page();
 545                 last_map_addr = phys_pud_init(pud, __pa(start), __pa(next),
 546                                                  page_size_mask);
 547
 548                 spin_lock(&init_mm.page_table_lock);
 549                 pgd_populate(&init_mm, pgd, pud);
 550                 spin_unlock(&init_mm.page_table_lock);
 551                 pgd_changed = true;
 552         }
 553
 554         if (pgd_changed)
 555                 sync_global_pgds(addr, end);
 556
 557         __flush_tlb_all();
 558
 559         return last_map_addr;
 560 }
 561
 562 #ifndef CONFIG_NUMA
 563 void __init initmem_init(void)
 564 {
 565         memblock_set_node(0, (phys_addr_t)ULLONG_MAX, 0);
 566 }
 567 #endif
 568
 569 void __init paging_init(void)
 570 {
 571         sparse_memory_present_with_active_regions(MAX_NUMNODES);
 572         sparse_init();
 573
 574         /*
 575          * clear the default setting with node 0
 576          * note: don't use nodes_clear here, that is really clearing when
 577          *       numa support is not compiled in, and later node_set_state
 578          *       will not set it back.
 579          */
 580         node_clear_state(0, N_MEMORY);
 581         if (N_MEMORY != N_NORMAL_MEMORY)
 582                 node_clear_state(0, N_NORMAL_MEMORY);
 583
 584         zone_sizes_init();
 585 }
 586
 587 /*
 588  * Memory hotplug specific functions
 589  */
 590 #ifdef CONFIG_MEMORY_HOTPLUG
 591 /*
 592  * After memory hotplug the variables max_pfn, max_low_pfn and high_memory need
 593  * updating.
 594  */
 595 static void  update_end_of_memory_vars(u64 start, u64 size)
 596 {
 597         unsigned long end_pfn = PFN_UP(start + size);
 598
 599         if (end_pfn > max_pfn) {
 600                 max_pfn = end_pfn;
 601                 max_low_pfn = end_pfn;
 602                 high_memory = (void *)__va(max_pfn * PAGE_SIZE - 1) + 1;
 603         }
 604 }
 605
 606 /*
 607  * Memory is added always to NORMAL zone. This means you will never get
 608  * additional DMA/DMA32 memory.
 609  */
 610 int arch_add_memory(int nid, u64 start, u64 size)
 611 {
 612         struct pglist_data *pgdat = NODE_DATA(nid);
 613         struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
 614         unsigned long start_pfn = start >> PAGE_SHIFT;
 615         unsigned long nr_pages = size >> PAGE_SHIFT;
 616         int ret;
 617
 618         init_memory_mapping(start, start + size);
 619
 620         ret = __add_pages(nid, zone, start_pfn, nr_pages);
 621         WARN_ON_ONCE(ret);
 622
 623         /* update max_pfn, max_low_pfn and high_memory */
 624         update_end_of_memory_vars(start, size);
 625
 626         return ret;
 627 }
 628 EXPORT_SYMBOL_GPL(arch_add_memory);
 629
 630 #endif /* CONFIG_MEMORY_HOTPLUG */
 631
 632 static struct kcore_list kcore_vsyscall;
 633
 634 static void __init register_page_bootmem_info(void)
 635 {
 636 #ifdef CONFIG_NUMA
 637         int i;
 638
 639         for_each_online_node(i)
 640                 register_page_bootmem_info_node(NODE_DATA(i));
 641 #endif
 642 }
 643
 644 void __init mem_init(void)
 645 {
 646         long codesize, reservedpages, datasize, initsize;
 647         unsigned long absent_pages;
 648
 649         pci_iommu_alloc();
 650
 651         /* clear_bss() already clear the empty_zero_page */
 652
 653         reservedpages = 0;
 654
 655         /* this will put all low memory onto the freelists */
 656         register_page_bootmem_info();
 657         totalram_pages = free_all_bootmem();
 658
 659         absent_pages = absent_pages_in_range(0, max_pfn);
 660         reservedpages = max_pfn - totalram_pages - absent_pages;
 661         after_bootmem = 1;
 662
 663         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
 664         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
 665         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
 666
 667         /* Register memory areas for /proc/kcore */
 668         kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
 669                          VSYSCALL_END - VSYSCALL_START, KCORE_OTHER);
 670
 671         printk(KERN_INFO "Memory: %luk/%luk available (%ldk kernel code, "
 672                          "%ldk absent, %ldk reserved, %ldk data, %ldk init)\n",
 673                 nr_free_pages() << (PAGE_SHIFT-10),
 674                 max_pfn << (PAGE_SHIFT-10),
 675                 codesize >> 10,
 676                 absent_pages << (PAGE_SHIFT-10),
 677                 reservedpages << (PAGE_SHIFT-10),
 678                 datasize >> 10,
 679                 initsize >> 10);
 680 }
 681
 682 #ifdef CONFIG_DEBUG_RODATA
 683 const int rodata_test_data = 0xC3;
 684 EXPORT_SYMBOL_GPL(rodata_test_data);
 685
 686 int kernel_set_to_readonly;
 687
 688 void set_kernel_text_rw(void)
 689 {
 690         unsigned long start = PFN_ALIGN(_text);
 691         unsigned long end = PFN_ALIGN(__stop___ex_table);
 692
 693         if (!kernel_set_to_readonly)
 694                 return;
 695
 696         pr_debug("Set kernel text: %lx - %lx for read write\n",
 697                  start, end);
 698
 699         /*
 700          * Make the kernel identity mapping for text RW. Kernel text
 701          * mapping will always be RO. Refer to the comment in
 702          * static_protections() in pageattr.c
 703          */
 704         set_memory_rw(start, (end - start) >> PAGE_SHIFT);
 705 }
 706
 707 void set_kernel_text_ro(void)
 708 {
 709         unsigned long start = PFN_ALIGN(_text);
 710         unsigned long end = PFN_ALIGN(__stop___ex_table);
 711
 712         if (!kernel_set_to_readonly)
 713                 return;
 714
 715         pr_debug("Set kernel text: %lx - %lx for read only\n",
 716                  start, end);
 717
 718         /*
 719          * Set the kernel identity mapping for text RO.
 720          */
 721         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
 722 }
 723
 724 void mark_rodata_ro(void)
 725 {
 726         unsigned long start = PFN_ALIGN(_text);
 727         unsigned long rodata_start =
 728                 ((unsigned long)__start_rodata + PAGE_SIZE - 1) & PAGE_MASK;
 729         unsigned long end = (unsigned long) &__end_rodata_hpage_align;
 730         unsigned long text_end = PAGE_ALIGN((unsigned long) &__stop___ex_table);
 731         unsigned long rodata_end = PAGE_ALIGN((unsigned long) &__end_rodata);
 732         unsigned long data_start = (unsigned long) &_sdata;
 733
 734         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
 735                (end - start) >> 10);
 736         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
 737
 738         kernel_set_to_readonly = 1;
 739
 740         /*
 741          * The rodata section (but not the kernel text!) should also be
 742          * not-executable.
 743          */
 744         set_memory_nx(rodata_start, (end - rodata_start) >> PAGE_SHIFT);
 745
 746         rodata_test();
 747
 748 #ifdef CONFIG_CPA_DEBUG
 749         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
 750         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
 751
 752         printk(KERN_INFO "Testing CPA: again\n");
 753         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
 754 #endif
 755
 756         free_init_pages("unused kernel memory",
 757                         (unsigned long) page_address(virt_to_page(text_end)),
 758                         (unsigned long)
 759                                  page_address(virt_to_page(rodata_start)));
 760         free_init_pages("unused kernel memory",
 761                         (unsigned long) page_address(virt_to_page(rodata_end)),
 762                         (unsigned long) page_address(virt_to_page(data_start)));
 763 }
 764
 765 #endif
 766
 767 int kern_addr_valid(unsigned long addr)
 768 {
 769         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
 770         pgd_t *pgd;
 771         pud_t *pud;
 772         pmd_t *pmd;
 773         pte_t *pte;
 774
 775         if (above != 0 && above != -1UL)
 776                 return 0;
 777
 778         pgd = pgd_offset_k(addr);
 779         if (pgd_none(*pgd))
 780                 return 0;
 781
 782         pud = pud_offset(pgd, addr);
 783         if (pud_none(*pud))
 784                 return 0;
 785
 786         pmd = pmd_offset(pud, addr);
 787         if (pmd_none(*pmd))
 788                 return 0;
 789
 790         if (pmd_large(*pmd))
 791                 return pfn_valid(pmd_pfn(*pmd));
 792
 793         pte = pte_offset_kernel(pmd, addr);
 794         if (pte_none(*pte))
 795                 return 0;
 796
 797         return pfn_valid(pte_pfn(*pte));
 798 }
 799
 800 /*
 801  * A pseudo VMA to allow ptrace access for the vsyscall page.  This only
 802  * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
 803  * not need special handling anymore:
 804  */
 805 static struct vm_area_struct gate_vma = {
 806         .vm_start       = VSYSCALL_START,
 807         .vm_end         = VSYSCALL_START + (VSYSCALL_MAPPED_PAGES * PAGE_SIZE),
 808         .vm_page_prot   = PAGE_READONLY_EXEC,
 809         .vm_flags       = VM_READ | VM_EXEC
 810 };
 811
 812 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
 813 {
 814 #ifdef CONFIG_IA32_EMULATION
 815         if (!mm || mm->context.ia32_compat)
 816                 return NULL;
 817 #endif
 818         return &gate_vma;
 819 }
 820
 821 int in_gate_area(struct mm_struct *mm, unsigned long addr)
 822 {
 823         struct vm_area_struct *vma = get_gate_vma(mm);
 824
 825         if (!vma)
 826                 return 0;
 827
 828         return (addr >= vma->vm_start) && (addr < vma->vm_end);
 829 }
 830
 831 /*
 832  * Use this when you have no reliable mm, typically from interrupt
 833  * context. It is less reliable than using a task's mm and may give
 834  * false positives.
 835  */
 836 int in_gate_area_no_mm(unsigned long addr)
 837 {
 838         return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);
 839 }
 840
 841 const char *arch_vma_name(struct vm_area_struct *vma)
 842 {
 843         if (vma->vm_mm && vma->vm_start == (long)vma->vm_mm->context.vdso)
 844                 return "[vdso]";
 845         if (vma == &gate_vma)
 846                 return "[vsyscall]";
 847         return NULL;
 848 }
 849
 850 #ifdef CONFIG_X86_UV
 851 unsigned long memory_block_size_bytes(void)
 852 {
 853         if (is_uv_system()) {
 854                 printk(KERN_INFO "UV: memory block size 2GB\n");
 855                 return 2UL * 1024 * 1024 * 1024;
 856         }
 857         return MIN_MEMORY_BLOCK_SIZE;
 858 }
 859 #endif
 860
 861 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 862 /*
 863  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
 864  */
 865 static long __meminitdata addr_start, addr_end;
 866 static void __meminitdata *p_start, *p_end;
 867 static int __meminitdata node_start;
 868
 869 int __meminit
 870 vmemmap_populate(struct page *start_page, unsigned long size, int node)
 871 {
 872         unsigned long addr = (unsigned long)start_page;
 873         unsigned long end = (unsigned long)(start_page + size);
 874         unsigned long next;
 875         pgd_t *pgd;
 876         pud_t *pud;
 877         pmd_t *pmd;
 878
 879         for (; addr < end; addr = next) {
 880                 void *p = NULL;
 881
 882                 pgd = vmemmap_pgd_populate(addr, node);
 883                 if (!pgd)
 884                         return -ENOMEM;
 885
 886                 pud = vmemmap_pud_populate(pgd, addr, node);
 887                 if (!pud)
 888                         return -ENOMEM;
 889
 890                 if (!cpu_has_pse) {
 891                         next = (addr + PAGE_SIZE) & PAGE_MASK;
 892                         pmd = vmemmap_pmd_populate(pud, addr, node);
 893
 894                         if (!pmd)
 895                                 return -ENOMEM;
 896
 897                         p = vmemmap_pte_populate(pmd, addr, node);
 898
 899                         if (!p)
 900                                 return -ENOMEM;
 901
 902                         addr_end = addr + PAGE_SIZE;
 903                         p_end = p + PAGE_SIZE;
 904                 } else {
 905                         next = pmd_addr_end(addr, end);
 906
 907                         pmd = pmd_offset(pud, addr);
 908                         if (pmd_none(*pmd)) {
 909                                 pte_t entry;
 910
 911                                 p = vmemmap_alloc_block_buf(PMD_SIZE, node);
 912                                 if (!p)
 913                                         return -ENOMEM;
 914
 915                                 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
 916                                                 PAGE_KERNEL_LARGE);
 917                                 set_pmd(pmd, __pmd(pte_val(entry)));
 918
 919                                 /* check to see if we have contiguous blocks */
 920                                 if (p_end != p || node_start != node) {
 921                                         if (p_start)
 922                                                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
 923                                                        addr_start, addr_end-1, p_start, p_end-1, node_start);
 924                                         addr_start = addr;
 925                                         node_start = node;
 926                                         p_start = p;
 927                                 }
 928
 929                                 addr_end = addr + PMD_SIZE;
 930                                 p_end = p + PMD_SIZE;
 931                         } else
 932                                 vmemmap_verify((pte_t *)pmd, node, addr, next);
 933                 }
 934
 935         }
 936         sync_global_pgds((unsigned long)start_page, end);
 937         return 0;
 938 }
 939
 940 void __meminit vmemmap_populate_print_last(void)
 941 {
 942         if (p_start) {
 943                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
 944                         addr_start, addr_end-1, p_start, p_end-1, node_start);
 945                 p_start = NULL;
 946                 p_end = NULL;
 947                 node_start = 0;
 948         }
 949 }
 950 #endif