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 void ident_pmd_init(unsigned long pmd_flag, pmd_t *pmd_page,
  60                            unsigned long addr, unsigned long end)
  61 {
  62         addr &= PMD_MASK;
  63         for (; addr < end; addr += PMD_SIZE) {
  64                 pmd_t *pmd = pmd_page + pmd_index(addr);
  65
  66                 if (!pmd_present(*pmd))
  67                         set_pmd(pmd, __pmd(addr | pmd_flag));
  68         }
  69 }
  70 static int ident_pud_init(struct x86_mapping_info *info, pud_t *pud_page,
  71                           unsigned long addr, unsigned long end)
  72 {
  73         unsigned long next;
  74
  75         for (; addr < end; addr = next) {
  76                 pud_t *pud = pud_page + pud_index(addr);
  77                 pmd_t *pmd;
  78
  79                 next = (addr & PUD_MASK) + PUD_SIZE;
  80                 if (next > end)
  81                         next = end;
  82
  83                 if (pud_present(*pud)) {
  84                         pmd = pmd_offset(pud, 0);
  85                         ident_pmd_init(info->pmd_flag, pmd, addr, next);
  86                         continue;
  87                 }
  88                 pmd = (pmd_t *)info->alloc_pgt_page(info->context);
  89                 if (!pmd)
  90                         return -ENOMEM;
  91                 ident_pmd_init(info->pmd_flag, pmd, addr, next);
  92                 set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE));
  93         }
  94
  95         return 0;
  96 }
  97
  98 int kernel_ident_mapping_init(struct x86_mapping_info *info, pgd_t *pgd_page,
  99                               unsigned long addr, unsigned long end)
 100 {
 101         unsigned long next;
 102         int result;
 103         int off = info->kernel_mapping ? pgd_index(__PAGE_OFFSET) : 0;
 104
 105         for (; addr < end; addr = next) {
 106                 pgd_t *pgd = pgd_page + pgd_index(addr) + off;
 107                 pud_t *pud;
 108
 109                 next = (addr & PGDIR_MASK) + PGDIR_SIZE;
 110                 if (next > end)
 111                         next = end;
 112
 113                 if (pgd_present(*pgd)) {
 114                         pud = pud_offset(pgd, 0);
 115                         result = ident_pud_init(info, pud, addr, next);
 116                         if (result)
 117                                 return result;
 118                         continue;
 119                 }
 120
 121                 pud = (pud_t *)info->alloc_pgt_page(info->context);
 122                 if (!pud)
 123                         return -ENOMEM;
 124                 result = ident_pud_init(info, pud, addr, next);
 125                 if (result)
 126                         return result;
 127                 set_pgd(pgd, __pgd(__pa(pud) | _KERNPG_TABLE));
 128         }
 129
 130         return 0;
 131 }
 132
 133 static int __init parse_direct_gbpages_off(char *arg)
 134 {
 135         direct_gbpages = 0;
 136         return 0;
 137 }
 138 early_param("nogbpages", parse_direct_gbpages_off);
 139
 140 static int __init parse_direct_gbpages_on(char *arg)
 141 {
 142         direct_gbpages = 1;
 143         return 0;
 144 }
 145 early_param("gbpages", parse_direct_gbpages_on);
 146
 147 /*
 148  * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
 149  * physical space so we can cache the place of the first one and move
 150  * around without checking the pgd every time.
 151  */
 152
 153 pteval_t __supported_pte_mask __read_mostly = ~_PAGE_IOMAP;
 154 EXPORT_SYMBOL_GPL(__supported_pte_mask);
 155
 156 int force_personality32;
 157
 158 /*
 159  * noexec32=on|off
 160  * Control non executable heap for 32bit processes.
 161  * To control the stack too use noexec=off
 162  *
 163  * on   PROT_READ does not imply PROT_EXEC for 32-bit processes (default)
 164  * off  PROT_READ implies PROT_EXEC
 165  */
 166 static int __init nonx32_setup(char *str)
 167 {
 168         if (!strcmp(str, "on"))
 169                 force_personality32 &= ~READ_IMPLIES_EXEC;
 170         else if (!strcmp(str, "off"))
 171                 force_personality32 |= READ_IMPLIES_EXEC;
 172         return 1;
 173 }
 174 __setup("noexec32=", nonx32_setup);
 175
 176 /*
 177  * When memory was added/removed make sure all the processes MM have
 178  * suitable PGD entries in the local PGD level page.
 179  */
 180 void sync_global_pgds(unsigned long start, unsigned long end)
 181 {
 182         unsigned long address;
 183
 184         for (address = start; address <= end; address += PGDIR_SIZE) {
 185                 const pgd_t *pgd_ref = pgd_offset_k(address);
 186                 struct page *page;
 187
 188                 if (pgd_none(*pgd_ref))
 189                         continue;
 190
 191                 spin_lock(&pgd_lock);
 192                 list_for_each_entry(page, &pgd_list, lru) {
 193                         pgd_t *pgd;
 194                         spinlock_t *pgt_lock;
 195
 196                         pgd = (pgd_t *)page_address(page) + pgd_index(address);
 197                         /* the pgt_lock only for Xen */
 198                         pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
 199                         spin_lock(pgt_lock);
 200
 201                         if (pgd_none(*pgd))
 202                                 set_pgd(pgd, *pgd_ref);
 203                         else
 204                                 BUG_ON(pgd_page_vaddr(*pgd)
 205                                        != pgd_page_vaddr(*pgd_ref));
 206
 207                         spin_unlock(pgt_lock);
 208                 }
 209                 spin_unlock(&pgd_lock);
 210         }
 211 }
 212
 213 /*
 214  * NOTE: This function is marked __ref because it calls __init function
 215  * (alloc_bootmem_pages). It's safe to do it ONLY when after_bootmem == 0.
 216  */
 217 static __ref void *spp_getpage(void)
 218 {
 219         void *ptr;
 220
 221         if (after_bootmem)
 222                 ptr = (void *) get_zeroed_page(GFP_ATOMIC | __GFP_NOTRACK);
 223         else
 224                 ptr = alloc_bootmem_pages(PAGE_SIZE);
 225
 226         if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
 227                 panic("set_pte_phys: cannot allocate page data %s\n",
 228                         after_bootmem ? "after bootmem" : "");
 229         }
 230
 231         pr_debug("spp_getpage %p\n", ptr);
 232
 233         return ptr;
 234 }
 235
 236 static pud_t *fill_pud(pgd_t *pgd, unsigned long vaddr)
 237 {
 238         if (pgd_none(*pgd)) {
 239                 pud_t *pud = (pud_t *)spp_getpage();
 240                 pgd_populate(&init_mm, pgd, pud);
 241                 if (pud != pud_offset(pgd, 0))
 242                         printk(KERN_ERR "PAGETABLE BUG #00! %p <-> %p\n",
 243                                pud, pud_offset(pgd, 0));
 244         }
 245         return pud_offset(pgd, vaddr);
 246 }
 247
 248 static pmd_t *fill_pmd(pud_t *pud, unsigned long vaddr)
 249 {
 250         if (pud_none(*pud)) {
 251                 pmd_t *pmd = (pmd_t *) spp_getpage();
 252                 pud_populate(&init_mm, pud, pmd);
 253                 if (pmd != pmd_offset(pud, 0))
 254                         printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
 255                                pmd, pmd_offset(pud, 0));
 256         }
 257         return pmd_offset(pud, vaddr);
 258 }
 259
 260 static pte_t *fill_pte(pmd_t *pmd, unsigned long vaddr)
 261 {
 262         if (pmd_none(*pmd)) {
 263                 pte_t *pte = (pte_t *) spp_getpage();
 264                 pmd_populate_kernel(&init_mm, pmd, pte);
 265                 if (pte != pte_offset_kernel(pmd, 0))
 266                         printk(KERN_ERR "PAGETABLE BUG #02!\n");
 267         }
 268         return pte_offset_kernel(pmd, vaddr);
 269 }
 270
 271 void set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
 272 {
 273         pud_t *pud;
 274         pmd_t *pmd;
 275         pte_t *pte;
 276
 277         pud = pud_page + pud_index(vaddr);
 278         pmd = fill_pmd(pud, vaddr);
 279         pte = fill_pte(pmd, vaddr);
 280
 281         set_pte(pte, new_pte);
 282
 283         /*
 284          * It's enough to flush this one mapping.
 285          * (PGE mappings get flushed as well)
 286          */
 287         __flush_tlb_one(vaddr);
 288 }
 289
 290 void set_pte_vaddr(unsigned long vaddr, pte_t pteval)
 291 {
 292         pgd_t *pgd;
 293         pud_t *pud_page;
 294
 295         pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
 296
 297         pgd = pgd_offset_k(vaddr);
 298         if (pgd_none(*pgd)) {
 299                 printk(KERN_ERR
 300                         "PGD FIXMAP MISSING, it should be setup in head.S!\n");
 301                 return;
 302         }
 303         pud_page = (pud_t*)pgd_page_vaddr(*pgd);
 304         set_pte_vaddr_pud(pud_page, vaddr, pteval);
 305 }
 306
 307 pmd_t * __init populate_extra_pmd(unsigned long vaddr)
 308 {
 309         pgd_t *pgd;
 310         pud_t *pud;
 311
 312         pgd = pgd_offset_k(vaddr);
 313         pud = fill_pud(pgd, vaddr);
 314         return fill_pmd(pud, vaddr);
 315 }
 316
 317 pte_t * __init populate_extra_pte(unsigned long vaddr)
 318 {
 319         pmd_t *pmd;
 320
 321         pmd = populate_extra_pmd(vaddr);
 322         return fill_pte(pmd, vaddr);
 323 }
 324
 325 /*
 326  * Create large page table mappings for a range of physical addresses.
 327  */
 328 static void __init __init_extra_mapping(unsigned long phys, unsigned long size,
 329                                                 pgprot_t prot)
 330 {
 331         pgd_t *pgd;
 332         pud_t *pud;
 333         pmd_t *pmd;
 334
 335         BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK));
 336         for (; size; phys += PMD_SIZE, size -= PMD_SIZE) {
 337                 pgd = pgd_offset_k((unsigned long)__va(phys));
 338                 if (pgd_none(*pgd)) {
 339                         pud = (pud_t *) spp_getpage();
 340                         set_pgd(pgd, __pgd(__pa(pud) | _KERNPG_TABLE |
 341                                                 _PAGE_USER));
 342                 }
 343                 pud = pud_offset(pgd, (unsigned long)__va(phys));
 344                 if (pud_none(*pud)) {
 345                         pmd = (pmd_t *) spp_getpage();
 346                         set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE |
 347                                                 _PAGE_USER));
 348                 }
 349                 pmd = pmd_offset(pud, phys);
 350                 BUG_ON(!pmd_none(*pmd));
 351                 set_pmd(pmd, __pmd(phys | pgprot_val(prot)));
 352         }
 353 }
 354
 355 void __init init_extra_mapping_wb(unsigned long phys, unsigned long size)
 356 {
 357         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE);
 358 }
 359
 360 void __init init_extra_mapping_uc(unsigned long phys, unsigned long size)
 361 {
 362         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE_NOCACHE);
 363 }
 364
 365 /*
 366  * The head.S code sets up the kernel high mapping:
 367  *
 368  *   from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
 369  *
 370  * phys_addr holds the negative offset to the kernel, which is added
 371  * to the compile time generated pmds. This results in invalid pmds up
 372  * to the point where we hit the physaddr 0 mapping.
 373  *
 374  * We limit the mappings to the region from _text to _brk_end.  _brk_end
 375  * is rounded up to the 2MB boundary. This catches the invalid pmds as
 376  * well, as they are located before _text:
 377  */
 378 void __init cleanup_highmap(void)
 379 {
 380         unsigned long vaddr = __START_KERNEL_map;
 381         unsigned long vaddr_end = __START_KERNEL_map + KERNEL_IMAGE_SIZE;
 382         unsigned long end = roundup((unsigned long)_brk_end, PMD_SIZE) - 1;
 383         pmd_t *pmd = level2_kernel_pgt;
 384
 385         /*
 386          * Native path, max_pfn_mapped is not set yet.
 387          * Xen has valid max_pfn_mapped set in
 388          *      arch/x86/xen/mmu.c:xen_setup_kernel_pagetable().
 389          */
 390         if (max_pfn_mapped)
 391                 vaddr_end = __START_KERNEL_map + (max_pfn_mapped << PAGE_SHIFT);
 392
 393         for (; vaddr + PMD_SIZE - 1 < vaddr_end; pmd++, vaddr += PMD_SIZE) {
 394                 if (pmd_none(*pmd))
 395                         continue;
 396                 if (vaddr < (unsigned long) _text || vaddr > end)
 397                         set_pmd(pmd, __pmd(0));
 398         }
 399 }
 400
 401 static unsigned long __meminit
 402 phys_pte_init(pte_t *pte_page, unsigned long addr, unsigned long end,
 403               pgprot_t prot)
 404 {
 405         unsigned long pages = 0, next;
 406         unsigned long last_map_addr = end;
 407         int i;
 408
 409         pte_t *pte = pte_page + pte_index(addr);
 410
 411         for (i = pte_index(addr); i < PTRS_PER_PTE; i++, addr = next, pte++) {
 412                 next = (addr & PAGE_MASK) + PAGE_SIZE;
 413                 if (addr >= end) {
 414                         if (!after_bootmem &&
 415                             !e820_any_mapped(addr & PAGE_MASK, next, E820_RAM) &&
 416                             !e820_any_mapped(addr & PAGE_MASK, next, E820_RESERVED_KERN))
 417                                 set_pte(pte, __pte(0));
 418                         continue;
 419                 }
 420
 421                 /*
 422                  * We will re-use the existing mapping.
 423                  * Xen for example has some special requirements, like mapping
 424                  * pagetable pages as RO. So assume someone who pre-setup
 425                  * these mappings are more intelligent.
 426                  */
 427                 if (pte_val(*pte)) {
 428                         if (!after_bootmem)
 429                                 pages++;
 430                         continue;
 431                 }
 432
 433                 if (0)
 434                         printk("   pte=%p addr=%lx pte=%016lx\n",
 435                                pte, addr, pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL).pte);
 436                 pages++;
 437                 set_pte(pte, pfn_pte(addr >> PAGE_SHIFT, prot));
 438                 last_map_addr = (addr & PAGE_MASK) + PAGE_SIZE;
 439         }
 440
 441         update_page_count(PG_LEVEL_4K, pages);
 442
 443         return last_map_addr;
 444 }
 445
 446 static unsigned long __meminit
 447 phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end,
 448               unsigned long page_size_mask, pgprot_t prot)
 449 {
 450         unsigned long pages = 0, next;
 451         unsigned long last_map_addr = end;
 452
 453         int i = pmd_index(address);
 454
 455         for (; i < PTRS_PER_PMD; i++, address = next) {
 456                 pmd_t *pmd = pmd_page + pmd_index(address);
 457                 pte_t *pte;
 458                 pgprot_t new_prot = prot;
 459
 460                 next = (address & PMD_MASK) + PMD_SIZE;
 461                 if (address >= end) {
 462                         if (!after_bootmem &&
 463                             !e820_any_mapped(address & PMD_MASK, next, E820_RAM) &&
 464                             !e820_any_mapped(address & PMD_MASK, next, E820_RESERVED_KERN))
 465                                 set_pmd(pmd, __pmd(0));
 466                         continue;
 467                 }
 468
 469                 if (pmd_val(*pmd)) {
 470                         if (!pmd_large(*pmd)) {
 471                                 spin_lock(&init_mm.page_table_lock);
 472                                 pte = (pte_t *)pmd_page_vaddr(*pmd);
 473                                 last_map_addr = phys_pte_init(pte, address,
 474                                                                 end, prot);
 475                                 spin_unlock(&init_mm.page_table_lock);
 476                                 continue;
 477                         }
 478                         /*
 479                          * If we are ok with PG_LEVEL_2M mapping, then we will
 480                          * use the existing mapping,
 481                          *
 482                          * Otherwise, we will split the large page mapping but
 483                          * use the same existing protection bits except for
 484                          * large page, so that we don't violate Intel's TLB
 485                          * Application note (317080) which says, while changing
 486                          * the page sizes, new and old translations should
 487                          * not differ with respect to page frame and
 488                          * attributes.
 489                          */
 490                         if (page_size_mask & (1 << PG_LEVEL_2M)) {
 491                                 if (!after_bootmem)
 492                                         pages++;
 493                                 last_map_addr = next;
 494                                 continue;
 495                         }
 496                         new_prot = pte_pgprot(pte_clrhuge(*(pte_t *)pmd));
 497                 }
 498
 499                 if (page_size_mask & (1<<PG_LEVEL_2M)) {
 500                         pages++;
 501                         spin_lock(&init_mm.page_table_lock);
 502                         set_pte((pte_t *)pmd,
 503                                 pfn_pte((address & PMD_MASK) >> PAGE_SHIFT,
 504                                         __pgprot(pgprot_val(prot) | _PAGE_PSE)));
 505                         spin_unlock(&init_mm.page_table_lock);
 506                         last_map_addr = next;
 507                         continue;
 508                 }
 509
 510                 pte = alloc_low_page();
 511                 last_map_addr = phys_pte_init(pte, address, end, new_prot);
 512
 513                 spin_lock(&init_mm.page_table_lock);
 514                 pmd_populate_kernel(&init_mm, pmd, pte);
 515                 spin_unlock(&init_mm.page_table_lock);
 516         }
 517         update_page_count(PG_LEVEL_2M, pages);
 518         return last_map_addr;
 519 }
 520
 521 static unsigned long __meminit
 522 phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end,
 523                          unsigned long page_size_mask)
 524 {
 525         unsigned long pages = 0, next;
 526         unsigned long last_map_addr = end;
 527         int i = pud_index(addr);
 528
 529         for (; i < PTRS_PER_PUD; i++, addr = next) {
 530                 pud_t *pud = pud_page + pud_index(addr);
 531                 pmd_t *pmd;
 532                 pgprot_t prot = PAGE_KERNEL;
 533
 534                 next = (addr & PUD_MASK) + PUD_SIZE;
 535                 if (addr >= end) {
 536                         if (!after_bootmem &&
 537                             !e820_any_mapped(addr & PUD_MASK, next, E820_RAM) &&
 538                             !e820_any_mapped(addr & PUD_MASK, next, E820_RESERVED_KERN))
 539                                 set_pud(pud, __pud(0));
 540                         continue;
 541                 }
 542
 543                 if (pud_val(*pud)) {
 544                         if (!pud_large(*pud)) {
 545                                 pmd = pmd_offset(pud, 0);
 546                                 last_map_addr = phys_pmd_init(pmd, addr, end,
 547                                                          page_size_mask, prot);
 548                                 __flush_tlb_all();
 549                                 continue;
 550                         }
 551                         /*
 552                          * If we are ok with PG_LEVEL_1G mapping, then we will
 553                          * use the existing mapping.
 554                          *
 555                          * Otherwise, we will split the gbpage mapping but use
 556                          * the same existing protection  bits except for large
 557                          * page, so that we don't violate Intel's TLB
 558                          * Application note (317080) which says, while changing
 559                          * the page sizes, new and old translations should
 560                          * not differ with respect to page frame and
 561                          * attributes.
 562                          */
 563                         if (page_size_mask & (1 << PG_LEVEL_1G)) {
 564                                 if (!after_bootmem)
 565                                         pages++;
 566                                 last_map_addr = next;
 567                                 continue;
 568                         }
 569                         prot = pte_pgprot(pte_clrhuge(*(pte_t *)pud));
 570                 }
 571
 572                 if (page_size_mask & (1<<PG_LEVEL_1G)) {
 573                         pages++;
 574                         spin_lock(&init_mm.page_table_lock);
 575                         set_pte((pte_t *)pud,
 576                                 pfn_pte((addr & PUD_MASK) >> PAGE_SHIFT,
 577                                         PAGE_KERNEL_LARGE));
 578                         spin_unlock(&init_mm.page_table_lock);
 579                         last_map_addr = next;
 580                         continue;
 581                 }
 582
 583                 pmd = alloc_low_page();
 584                 last_map_addr = phys_pmd_init(pmd, addr, end, page_size_mask,
 585                                               prot);
 586
 587                 spin_lock(&init_mm.page_table_lock);
 588                 pud_populate(&init_mm, pud, pmd);
 589                 spin_unlock(&init_mm.page_table_lock);
 590         }
 591         __flush_tlb_all();
 592
 593         update_page_count(PG_LEVEL_1G, pages);
 594
 595         return last_map_addr;
 596 }
 597
 598 unsigned long __meminit
 599 kernel_physical_mapping_init(unsigned long start,
 600                              unsigned long end,
 601                              unsigned long page_size_mask)
 602 {
 603         bool pgd_changed = false;
 604         unsigned long next, last_map_addr = end;
 605         unsigned long addr;
 606
 607         start = (unsigned long)__va(start);
 608         end = (unsigned long)__va(end);
 609         addr = start;
 610
 611         for (; start < end; start = next) {
 612                 pgd_t *pgd = pgd_offset_k(start);
 613                 pud_t *pud;
 614
 615                 next = (start & PGDIR_MASK) + PGDIR_SIZE;
 616
 617                 if (pgd_val(*pgd)) {
 618                         pud = (pud_t *)pgd_page_vaddr(*pgd);
 619                         last_map_addr = phys_pud_init(pud, __pa(start),
 620                                                  __pa(end), page_size_mask);
 621                         continue;
 622                 }
 623
 624                 pud = alloc_low_page();
 625                 last_map_addr = phys_pud_init(pud, __pa(start), __pa(end),
 626                                                  page_size_mask);
 627
 628                 spin_lock(&init_mm.page_table_lock);
 629                 pgd_populate(&init_mm, pgd, pud);
 630                 spin_unlock(&init_mm.page_table_lock);
 631                 pgd_changed = true;
 632         }
 633
 634         if (pgd_changed)
 635                 sync_global_pgds(addr, end - 1);
 636
 637         __flush_tlb_all();
 638
 639         return last_map_addr;
 640 }
 641
 642 #ifndef CONFIG_NUMA
 643 void __init initmem_init(void)
 644 {
 645         memblock_set_node(0, (phys_addr_t)ULLONG_MAX, 0);
 646 }
 647 #endif
 648
 649 void __init paging_init(void)
 650 {
 651         sparse_memory_present_with_active_regions(MAX_NUMNODES);
 652         sparse_init();
 653
 654         /*
 655          * clear the default setting with node 0
 656          * note: don't use nodes_clear here, that is really clearing when
 657          *       numa support is not compiled in, and later node_set_state
 658          *       will not set it back.
 659          */
 660         node_clear_state(0, N_MEMORY);
 661         if (N_MEMORY != N_NORMAL_MEMORY)
 662                 node_clear_state(0, N_NORMAL_MEMORY);
 663
 664         zone_sizes_init();
 665 }
 666
 667 /*
 668  * Memory hotplug specific functions
 669  */
 670 #ifdef CONFIG_MEMORY_HOTPLUG
 671 /*
 672  * After memory hotplug the variables max_pfn, max_low_pfn and high_memory need
 673  * updating.
 674  */
 675 static void  update_end_of_memory_vars(u64 start, u64 size)
 676 {
 677         unsigned long end_pfn = PFN_UP(start + size);
 678
 679         if (end_pfn > max_pfn) {
 680                 max_pfn = end_pfn;
 681                 max_low_pfn = end_pfn;
 682                 high_memory = (void *)__va(max_pfn * PAGE_SIZE - 1) + 1;
 683         }
 684 }
 685
 686 /*
 687  * Memory is added always to NORMAL zone. This means you will never get
 688  * additional DMA/DMA32 memory.
 689  */
 690 int arch_add_memory(int nid, u64 start, u64 size)
 691 {
 692         struct pglist_data *pgdat = NODE_DATA(nid);
 693         struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
 694         unsigned long start_pfn = start >> PAGE_SHIFT;
 695         unsigned long nr_pages = size >> PAGE_SHIFT;
 696         int ret;
 697
 698         init_memory_mapping(start, start + size);
 699
 700         ret = __add_pages(nid, zone, start_pfn, nr_pages);
 701         WARN_ON_ONCE(ret);
 702
 703         /* update max_pfn, max_low_pfn and high_memory */
 704         update_end_of_memory_vars(start, size);
 705
 706         return ret;
 707 }
 708 EXPORT_SYMBOL_GPL(arch_add_memory);
 709
 710 #ifdef CONFIG_MEMORY_HOTREMOVE
 711 int __ref arch_remove_memory(u64 start, u64 size)
 712 {
 713         unsigned long start_pfn = start >> PAGE_SHIFT;
 714         unsigned long nr_pages = size >> PAGE_SHIFT;
 715         struct zone *zone;
 716         int ret;
 717
 718         zone = page_zone(pfn_to_page(start_pfn));
 719         ret = __remove_pages(zone, start_pfn, nr_pages);
 720         WARN_ON_ONCE(ret);
 721
 722         return ret;
 723 }
 724 #endif
 725 #endif /* CONFIG_MEMORY_HOTPLUG */
 726
 727 static struct kcore_list kcore_vsyscall;
 728
 729 static void __init register_page_bootmem_info(void)
 730 {
 731 #ifdef CONFIG_NUMA
 732         int i;
 733
 734         for_each_online_node(i)
 735                 register_page_bootmem_info_node(NODE_DATA(i));
 736 #endif
 737 }
 738
 739 void __init mem_init(void)
 740 {
 741         long codesize, reservedpages, datasize, initsize;
 742         unsigned long absent_pages;
 743
 744         pci_iommu_alloc();
 745
 746         /* clear_bss() already clear the empty_zero_page */
 747
 748         reservedpages = 0;
 749
 750         /* this will put all low memory onto the freelists */
 751         register_page_bootmem_info();
 752         totalram_pages = free_all_bootmem();
 753
 754         absent_pages = absent_pages_in_range(0, max_pfn);
 755         reservedpages = max_pfn - totalram_pages - absent_pages;
 756         after_bootmem = 1;
 757
 758         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
 759         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
 760         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
 761
 762         /* Register memory areas for /proc/kcore */
 763         kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
 764                          VSYSCALL_END - VSYSCALL_START, KCORE_OTHER);
 765
 766         printk(KERN_INFO "Memory: %luk/%luk available (%ldk kernel code, "
 767                          "%ldk absent, %ldk reserved, %ldk data, %ldk init)\n",
 768                 nr_free_pages() << (PAGE_SHIFT-10),
 769                 max_pfn << (PAGE_SHIFT-10),
 770                 codesize >> 10,
 771                 absent_pages << (PAGE_SHIFT-10),
 772                 reservedpages << (PAGE_SHIFT-10),
 773                 datasize >> 10,
 774                 initsize >> 10);
 775 }
 776
 777 #ifdef CONFIG_DEBUG_RODATA
 778 const int rodata_test_data = 0xC3;
 779 EXPORT_SYMBOL_GPL(rodata_test_data);
 780
 781 int kernel_set_to_readonly;
 782
 783 void set_kernel_text_rw(void)
 784 {
 785         unsigned long start = PFN_ALIGN(_text);
 786         unsigned long end = PFN_ALIGN(__stop___ex_table);
 787
 788         if (!kernel_set_to_readonly)
 789                 return;
 790
 791         pr_debug("Set kernel text: %lx - %lx for read write\n",
 792                  start, end);
 793
 794         /*
 795          * Make the kernel identity mapping for text RW. Kernel text
 796          * mapping will always be RO. Refer to the comment in
 797          * static_protections() in pageattr.c
 798          */
 799         set_memory_rw(start, (end - start) >> PAGE_SHIFT);
 800 }
 801
 802 void set_kernel_text_ro(void)
 803 {
 804         unsigned long start = PFN_ALIGN(_text);
 805         unsigned long end = PFN_ALIGN(__stop___ex_table);
 806
 807         if (!kernel_set_to_readonly)
 808                 return;
 809
 810         pr_debug("Set kernel text: %lx - %lx for read only\n",
 811                  start, end);
 812
 813         /*
 814          * Set the kernel identity mapping for text RO.
 815          */
 816         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
 817 }
 818
 819 void mark_rodata_ro(void)
 820 {
 821         unsigned long start = PFN_ALIGN(_text);
 822         unsigned long rodata_start = PFN_ALIGN(__start_rodata);
 823         unsigned long end = (unsigned long) &__end_rodata_hpage_align;
 824         unsigned long text_end = PFN_ALIGN(&__stop___ex_table);
 825         unsigned long rodata_end = PFN_ALIGN(&__end_rodata);
 826         unsigned long all_end = PFN_ALIGN(&_end);
 827
 828         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
 829                (end - start) >> 10);
 830         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
 831
 832         kernel_set_to_readonly = 1;
 833
 834         /*
 835          * The rodata/data/bss/brk section (but not the kernel text!)
 836          * should also be not-executable.
 837          */
 838         set_memory_nx(rodata_start, (all_end - rodata_start) >> PAGE_SHIFT);
 839
 840         rodata_test();
 841
 842 #ifdef CONFIG_CPA_DEBUG
 843         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
 844         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
 845
 846         printk(KERN_INFO "Testing CPA: again\n");
 847         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
 848 #endif
 849
 850         free_init_pages("unused kernel memory",
 851                         (unsigned long) __va(__pa_symbol(text_end)),
 852                         (unsigned long) __va(__pa_symbol(rodata_start)));
 853
 854         free_init_pages("unused kernel memory",
 855                         (unsigned long) __va(__pa_symbol(rodata_end)),
 856                         (unsigned long) __va(__pa_symbol(_sdata)));
 857 }
 858
 859 #endif
 860
 861 int kern_addr_valid(unsigned long addr)
 862 {
 863         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
 864         pgd_t *pgd;
 865         pud_t *pud;
 866         pmd_t *pmd;
 867         pte_t *pte;
 868
 869         if (above != 0 && above != -1UL)
 870                 return 0;
 871
 872         pgd = pgd_offset_k(addr);
 873         if (pgd_none(*pgd))
 874                 return 0;
 875
 876         pud = pud_offset(pgd, addr);
 877         if (pud_none(*pud))
 878                 return 0;
 879
 880         if (pud_large(*pud))
 881                 return pfn_valid(pud_pfn(*pud));
 882
 883         pmd = pmd_offset(pud, addr);
 884         if (pmd_none(*pmd))
 885                 return 0;
 886
 887         if (pmd_large(*pmd))
 888                 return pfn_valid(pmd_pfn(*pmd));
 889
 890         pte = pte_offset_kernel(pmd, addr);
 891         if (pte_none(*pte))
 892                 return 0;
 893
 894         return pfn_valid(pte_pfn(*pte));
 895 }
 896
 897 /*
 898  * A pseudo VMA to allow ptrace access for the vsyscall page.  This only
 899  * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
 900  * not need special handling anymore:
 901  */
 902 static struct vm_area_struct gate_vma = {
 903         .vm_start       = VSYSCALL_START,
 904         .vm_end         = VSYSCALL_START + (VSYSCALL_MAPPED_PAGES * PAGE_SIZE),
 905         .vm_page_prot   = PAGE_READONLY_EXEC,
 906         .vm_flags       = VM_READ | VM_EXEC
 907 };
 908
 909 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
 910 {
 911 #ifdef CONFIG_IA32_EMULATION
 912         if (!mm || mm->context.ia32_compat)
 913                 return NULL;
 914 #endif
 915         return &gate_vma;
 916 }
 917
 918 int in_gate_area(struct mm_struct *mm, unsigned long addr)
 919 {
 920         struct vm_area_struct *vma = get_gate_vma(mm);
 921
 922         if (!vma)
 923                 return 0;
 924
 925         return (addr >= vma->vm_start) && (addr < vma->vm_end);
 926 }
 927
 928 /*
 929  * Use this when you have no reliable mm, typically from interrupt
 930  * context. It is less reliable than using a task's mm and may give
 931  * false positives.
 932  */
 933 int in_gate_area_no_mm(unsigned long addr)
 934 {
 935         return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);
 936 }
 937
 938 const char *arch_vma_name(struct vm_area_struct *vma)
 939 {
 940         if (vma->vm_mm && vma->vm_start == (long)vma->vm_mm->context.vdso)
 941                 return "[vdso]";
 942         if (vma == &gate_vma)
 943                 return "[vsyscall]";
 944         return NULL;
 945 }
 946
 947 #ifdef CONFIG_X86_UV
 948 unsigned long memory_block_size_bytes(void)
 949 {
 950         if (is_uv_system()) {
 951                 printk(KERN_INFO "UV: memory block size 2GB\n");
 952                 return 2UL * 1024 * 1024 * 1024;
 953         }
 954         return MIN_MEMORY_BLOCK_SIZE;
 955 }
 956 #endif
 957
 958 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 959 /*
 960  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
 961  */
 962 static long __meminitdata addr_start, addr_end;
 963 static void __meminitdata *p_start, *p_end;
 964 static int __meminitdata node_start;
 965
 966 int __meminit
 967 vmemmap_populate(struct page *start_page, unsigned long size, int node)
 968 {
 969         unsigned long addr = (unsigned long)start_page;
 970         unsigned long end = (unsigned long)(start_page + size);
 971         unsigned long next;
 972         pgd_t *pgd;
 973         pud_t *pud;
 974         pmd_t *pmd;
 975
 976         for (; addr < end; addr = next) {
 977                 void *p = NULL;
 978
 979                 pgd = vmemmap_pgd_populate(addr, node);
 980                 if (!pgd)
 981                         return -ENOMEM;
 982
 983                 pud = vmemmap_pud_populate(pgd, addr, node);
 984                 if (!pud)
 985                         return -ENOMEM;
 986
 987                 if (!cpu_has_pse) {
 988                         next = (addr + PAGE_SIZE) & PAGE_MASK;
 989                         pmd = vmemmap_pmd_populate(pud, addr, node);
 990
 991                         if (!pmd)
 992                                 return -ENOMEM;
 993
 994                         p = vmemmap_pte_populate(pmd, addr, node);
 995
 996                         if (!p)
 997                                 return -ENOMEM;
 998
 999                         addr_end = addr + PAGE_SIZE;
1000                         p_end = p + PAGE_SIZE;
1001                 } else {
1002                         next = pmd_addr_end(addr, end);
1003
1004                         pmd = pmd_offset(pud, addr);
1005                         if (pmd_none(*pmd)) {
1006                                 pte_t entry;
1007
1008                                 p = vmemmap_alloc_block_buf(PMD_SIZE, node);
1009                                 if (!p)
1010                                         return -ENOMEM;
1011
1012                                 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1013                                                 PAGE_KERNEL_LARGE);
1014                                 set_pmd(pmd, __pmd(pte_val(entry)));
1015
1016                                 /* check to see if we have contiguous blocks */
1017                                 if (p_end != p || node_start != node) {
1018                                         if (p_start)
1019                                                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1020                                                        addr_start, addr_end-1, p_start, p_end-1, node_start);
1021                                         addr_start = addr;
1022                                         node_start = node;
1023                                         p_start = p;
1024                                 }
1025
1026                                 addr_end = addr + PMD_SIZE;
1027                                 p_end = p + PMD_SIZE;
1028                         } else
1029                                 vmemmap_verify((pte_t *)pmd, node, addr, next);
1030                 }
1031
1032         }
1033         sync_global_pgds((unsigned long)start_page, end - 1);
1034         return 0;
1035 }
1036
1037 #if defined(CONFIG_MEMORY_HOTPLUG_SPARSE) && defined(CONFIG_HAVE_BOOTMEM_INFO_NODE)
1038 void register_page_bootmem_memmap(unsigned long section_nr,
1039                                   struct page *start_page, unsigned long size)
1040 {
1041         unsigned long addr = (unsigned long)start_page;
1042         unsigned long end = (unsigned long)(start_page + size);
1043         unsigned long next;
1044         pgd_t *pgd;
1045         pud_t *pud;
1046         pmd_t *pmd;
1047         unsigned int nr_pages;
1048         struct page *page;
1049
1050         for (; addr < end; addr = next) {
1051                 pte_t *pte = NULL;
1052
1053                 pgd = pgd_offset_k(addr);
1054                 if (pgd_none(*pgd)) {
1055                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1056                         continue;
1057                 }
1058                 get_page_bootmem(section_nr, pgd_page(*pgd), MIX_SECTION_INFO);
1059
1060                 pud = pud_offset(pgd, addr);
1061                 if (pud_none(*pud)) {
1062                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1063                         continue;
1064                 }
1065                 get_page_bootmem(section_nr, pud_page(*pud), MIX_SECTION_INFO);
1066
1067                 if (!cpu_has_pse) {
1068                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1069                         pmd = pmd_offset(pud, addr);
1070                         if (pmd_none(*pmd))
1071                                 continue;
1072                         get_page_bootmem(section_nr, pmd_page(*pmd),
1073                                          MIX_SECTION_INFO);
1074
1075                         pte = pte_offset_kernel(pmd, addr);
1076                         if (pte_none(*pte))
1077                                 continue;
1078                         get_page_bootmem(section_nr, pte_page(*pte),
1079                                          SECTION_INFO);
1080                 } else {
1081                         next = pmd_addr_end(addr, end);
1082
1083                         pmd = pmd_offset(pud, addr);
1084                         if (pmd_none(*pmd))
1085                                 continue;
1086
1087                         nr_pages = 1 << (get_order(PMD_SIZE));
1088                         page = pmd_page(*pmd);
1089                         while (nr_pages--)
1090                                 get_page_bootmem(section_nr, page++,
1091                                                  SECTION_INFO);
1092                 }
1093         }
1094 }
1095 #endif
1096
1097 void __meminit vmemmap_populate_print_last(void)
1098 {
1099         if (p_start) {
1100                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1101                         addr_start, addr_end-1, p_start, p_end-1, node_start);
1102                 p_start = NULL;
1103                 p_end = NULL;
1104                 node_start = 0;
1105         }
1106 }
1107 #endif