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/proc_fs.h>
  25 #include <linux/pci.h>
  26 #include <linux/pfn.h>
  27 #include <linux/poison.h>
  28 #include <linux/dma-mapping.h>
  29 #include <linux/module.h>
  30 #include <linux/memory_hotplug.h>
  31 #include <linux/nmi.h>
  32 #include <linux/gfp.h>
  33
  34 #include <asm/processor.h>
  35 #include <asm/bios_ebda.h>
  36 #include <asm/system.h>
  37 #include <asm/uaccess.h>
  38 #include <asm/pgtable.h>
  39 #include <asm/pgalloc.h>
  40 #include <asm/dma.h>
  41 #include <asm/fixmap.h>
  42 #include <asm/e820.h>
  43 #include <asm/apic.h>
  44 #include <asm/tlb.h>
  45 #include <asm/mmu_context.h>
  46 #include <asm/proto.h>
  47 #include <asm/smp.h>
  48 #include <asm/sections.h>
  49 #include <asm/kdebug.h>
  50 #include <asm/numa.h>
  51 #include <asm/cacheflush.h>
  52 #include <asm/init.h>
  53 #include <linux/bootmem.h>
  54
  55 static unsigned long dma_reserve __initdata;
  56
  57 static int __init parse_direct_gbpages_off(char *arg)
  58 {
  59         direct_gbpages = 0;
  60         return 0;
  61 }
  62 early_param("nogbpages", parse_direct_gbpages_off);
  63
  64 static int __init parse_direct_gbpages_on(char *arg)
  65 {
  66         direct_gbpages = 1;
  67         return 0;
  68 }
  69 early_param("gbpages", parse_direct_gbpages_on);
  70
  71 /*
  72  * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
  73  * physical space so we can cache the place of the first one and move
  74  * around without checking the pgd every time.
  75  */
  76
  77 pteval_t __supported_pte_mask __read_mostly = ~_PAGE_IOMAP;
  78 EXPORT_SYMBOL_GPL(__supported_pte_mask);
  79
  80 int force_personality32;
  81
  82 /*
  83  * noexec32=on|off
  84  * Control non executable heap for 32bit processes.
  85  * To control the stack too use noexec=off
  86  *
  87  * on   PROT_READ does not imply PROT_EXEC for 32-bit processes (default)
  88  * off  PROT_READ implies PROT_EXEC
  89  */
  90 static int __init nonx32_setup(char *str)
  91 {
  92         if (!strcmp(str, "on"))
  93                 force_personality32 &= ~READ_IMPLIES_EXEC;
  94         else if (!strcmp(str, "off"))
  95                 force_personality32 |= READ_IMPLIES_EXEC;
  96         return 1;
  97 }
  98 __setup("noexec32=", nonx32_setup);
  99
 100 /*
 101  * When memory was added/removed make sure all the processes MM have
 102  * suitable PGD entries in the local PGD level page.
 103  */
 104 void sync_global_pgds(unsigned long start, unsigned long end)
 105 {
 106         unsigned long address;
 107
 108         for (address = start; address <= end; address += PGDIR_SIZE) {
 109                 const pgd_t *pgd_ref = pgd_offset_k(address);
 110                 unsigned long flags;
 111                 struct page *page;
 112
 113                 if (pgd_none(*pgd_ref))
 114                         continue;
 115
 116                 spin_lock_irqsave(&pgd_lock, flags);
 117                 list_for_each_entry(page, &pgd_list, lru) {
 118                         pgd_t *pgd;
 119                         spinlock_t *pgt_lock;
 120
 121                         pgd = (pgd_t *)page_address(page) + pgd_index(address);
 122                         pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
 123                         spin_lock(pgt_lock);
 124
 125                         if (pgd_none(*pgd))
 126                                 set_pgd(pgd, *pgd_ref);
 127                         else
 128                                 BUG_ON(pgd_page_vaddr(*pgd)
 129                                        != pgd_page_vaddr(*pgd_ref));
 130
 131                         spin_unlock(pgt_lock);
 132                 }
 133                 spin_unlock_irqrestore(&pgd_lock, flags);
 134         }
 135 }
 136
 137 /*
 138  * NOTE: This function is marked __ref because it calls __init function
 139  * (alloc_bootmem_pages). It's safe to do it ONLY when after_bootmem == 0.
 140  */
 141 static __ref void *spp_getpage(void)
 142 {
 143         void *ptr;
 144
 145         if (after_bootmem)
 146                 ptr = (void *) get_zeroed_page(GFP_ATOMIC | __GFP_NOTRACK);
 147         else
 148                 ptr = alloc_bootmem_pages(PAGE_SIZE);
 149
 150         if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
 151                 panic("set_pte_phys: cannot allocate page data %s\n",
 152                         after_bootmem ? "after bootmem" : "");
 153         }
 154
 155         pr_debug("spp_getpage %p\n", ptr);
 156
 157         return ptr;
 158 }
 159
 160 static pud_t *fill_pud(pgd_t *pgd, unsigned long vaddr)
 161 {
 162         if (pgd_none(*pgd)) {
 163                 pud_t *pud = (pud_t *)spp_getpage();
 164                 pgd_populate(&init_mm, pgd, pud);
 165                 if (pud != pud_offset(pgd, 0))
 166                         printk(KERN_ERR "PAGETABLE BUG #00! %p <-> %p\n",
 167                                pud, pud_offset(pgd, 0));
 168         }
 169         return pud_offset(pgd, vaddr);
 170 }
 171
 172 static pmd_t *fill_pmd(pud_t *pud, unsigned long vaddr)
 173 {
 174         if (pud_none(*pud)) {
 175                 pmd_t *pmd = (pmd_t *) spp_getpage();
 176                 pud_populate(&init_mm, pud, pmd);
 177                 if (pmd != pmd_offset(pud, 0))
 178                         printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
 179                                pmd, pmd_offset(pud, 0));
 180         }
 181         return pmd_offset(pud, vaddr);
 182 }
 183
 184 static pte_t *fill_pte(pmd_t *pmd, unsigned long vaddr)
 185 {
 186         if (pmd_none(*pmd)) {
 187                 pte_t *pte = (pte_t *) spp_getpage();
 188                 pmd_populate_kernel(&init_mm, pmd, pte);
 189                 if (pte != pte_offset_kernel(pmd, 0))
 190                         printk(KERN_ERR "PAGETABLE BUG #02!\n");
 191         }
 192         return pte_offset_kernel(pmd, vaddr);
 193 }
 194
 195 void set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
 196 {
 197         pud_t *pud;
 198         pmd_t *pmd;
 199         pte_t *pte;
 200
 201         pud = pud_page + pud_index(vaddr);
 202         pmd = fill_pmd(pud, vaddr);
 203         pte = fill_pte(pmd, vaddr);
 204
 205         set_pte(pte, new_pte);
 206
 207         /*
 208          * It's enough to flush this one mapping.
 209          * (PGE mappings get flushed as well)
 210          */
 211         __flush_tlb_one(vaddr);
 212 }
 213
 214 void set_pte_vaddr(unsigned long vaddr, pte_t pteval)
 215 {
 216         pgd_t *pgd;
 217         pud_t *pud_page;
 218
 219         pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
 220
 221         pgd = pgd_offset_k(vaddr);
 222         if (pgd_none(*pgd)) {
 223                 printk(KERN_ERR
 224                         "PGD FIXMAP MISSING, it should be setup in head.S!\n");
 225                 return;
 226         }
 227         pud_page = (pud_t*)pgd_page_vaddr(*pgd);
 228         set_pte_vaddr_pud(pud_page, vaddr, pteval);
 229 }
 230
 231 pmd_t * __init populate_extra_pmd(unsigned long vaddr)
 232 {
 233         pgd_t *pgd;
 234         pud_t *pud;
 235
 236         pgd = pgd_offset_k(vaddr);
 237         pud = fill_pud(pgd, vaddr);
 238         return fill_pmd(pud, vaddr);
 239 }
 240
 241 pte_t * __init populate_extra_pte(unsigned long vaddr)
 242 {
 243         pmd_t *pmd;
 244
 245         pmd = populate_extra_pmd(vaddr);
 246         return fill_pte(pmd, vaddr);
 247 }
 248
 249 /*
 250  * Create large page table mappings for a range of physical addresses.
 251  */
 252 static void __init __init_extra_mapping(unsigned long phys, unsigned long size,
 253                                                 pgprot_t prot)
 254 {
 255         pgd_t *pgd;
 256         pud_t *pud;
 257         pmd_t *pmd;
 258
 259         BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK));
 260         for (; size; phys += PMD_SIZE, size -= PMD_SIZE) {
 261                 pgd = pgd_offset_k((unsigned long)__va(phys));
 262                 if (pgd_none(*pgd)) {
 263                         pud = (pud_t *) spp_getpage();
 264                         set_pgd(pgd, __pgd(__pa(pud) | _KERNPG_TABLE |
 265                                                 _PAGE_USER));
 266                 }
 267                 pud = pud_offset(pgd, (unsigned long)__va(phys));
 268                 if (pud_none(*pud)) {
 269                         pmd = (pmd_t *) spp_getpage();
 270                         set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE |
 271                                                 _PAGE_USER));
 272                 }
 273                 pmd = pmd_offset(pud, phys);
 274                 BUG_ON(!pmd_none(*pmd));
 275                 set_pmd(pmd, __pmd(phys | pgprot_val(prot)));
 276         }
 277 }
 278
 279 void __init init_extra_mapping_wb(unsigned long phys, unsigned long size)
 280 {
 281         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE);
 282 }
 283
 284 void __init init_extra_mapping_uc(unsigned long phys, unsigned long size)
 285 {
 286         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE_NOCACHE);
 287 }
 288
 289 /*
 290  * The head.S code sets up the kernel high mapping:
 291  *
 292  *   from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
 293  *
 294  * phys_addr holds the negative offset to the kernel, which is added
 295  * to the compile time generated pmds. This results in invalid pmds up
 296  * to the point where we hit the physaddr 0 mapping.
 297  *
 298  * We limit the mappings to the region from _text to _end.  _end is
 299  * rounded up to the 2MB boundary. This catches the invalid pmds as
 300  * well, as they are located before _text:
 301  */
 302 void __init cleanup_highmap(void)
 303 {
 304         unsigned long vaddr = __START_KERNEL_map;
 305         unsigned long end = roundup((unsigned long)_end, PMD_SIZE) - 1;
 306         pmd_t *pmd = level2_kernel_pgt;
 307         pmd_t *last_pmd = pmd + PTRS_PER_PMD;
 308
 309         for (; pmd < last_pmd; pmd++, vaddr += PMD_SIZE) {
 310                 if (pmd_none(*pmd))
 311                         continue;
 312                 if (vaddr < (unsigned long) _text || vaddr > end)
 313                         set_pmd(pmd, __pmd(0));
 314         }
 315 }
 316
 317 static __ref void *alloc_low_page(unsigned long *phys)
 318 {
 319         unsigned long pfn = e820_table_end++;
 320         void *adr;
 321
 322         if (after_bootmem) {
 323                 adr = (void *)get_zeroed_page(GFP_ATOMIC | __GFP_NOTRACK);
 324                 *phys = __pa(adr);
 325
 326                 return adr;
 327         }
 328
 329         if (pfn >= e820_table_top)
 330                 panic("alloc_low_page: ran out of memory");
 331
 332         adr = early_memremap(pfn * PAGE_SIZE, PAGE_SIZE);
 333         clear_page(adr);
 334         *phys  = pfn * PAGE_SIZE;
 335         return adr;
 336 }
 337
 338 static __ref void unmap_low_page(void *adr)
 339 {
 340         if (after_bootmem)
 341                 return;
 342
 343         early_iounmap(adr, PAGE_SIZE);
 344 }
 345
 346 static unsigned long __meminit
 347 phys_pte_init(pte_t *pte_page, unsigned long addr, unsigned long end,
 348               pgprot_t prot)
 349 {
 350         unsigned pages = 0;
 351         unsigned long last_map_addr = end;
 352         int i;
 353
 354         pte_t *pte = pte_page + pte_index(addr);
 355
 356         for(i = pte_index(addr); i < PTRS_PER_PTE; i++, addr += PAGE_SIZE, pte++) {
 357
 358                 if (addr >= end) {
 359                         if (!after_bootmem) {
 360                                 for(; i < PTRS_PER_PTE; i++, pte++)
 361                                         set_pte(pte, __pte(0));
 362                         }
 363                         break;
 364                 }
 365
 366                 /*
 367                  * We will re-use the existing mapping.
 368                  * Xen for example has some special requirements, like mapping
 369                  * pagetable pages as RO. So assume someone who pre-setup
 370                  * these mappings are more intelligent.
 371                  */
 372                 if (pte_val(*pte)) {
 373                         pages++;
 374                         continue;
 375                 }
 376
 377                 if (0)
 378                         printk("   pte=%p addr=%lx pte=%016lx\n",
 379                                pte, addr, pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL).pte);
 380                 pages++;
 381                 set_pte(pte, pfn_pte(addr >> PAGE_SHIFT, prot));
 382                 last_map_addr = (addr & PAGE_MASK) + PAGE_SIZE;
 383         }
 384
 385         update_page_count(PG_LEVEL_4K, pages);
 386
 387         return last_map_addr;
 388 }
 389
 390 static unsigned long __meminit
 391 phys_pte_update(pmd_t *pmd, unsigned long address, unsigned long end,
 392                 pgprot_t prot)
 393 {
 394         pte_t *pte = (pte_t *)pmd_page_vaddr(*pmd);
 395
 396         return phys_pte_init(pte, address, end, prot);
 397 }
 398
 399 static unsigned long __meminit
 400 phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end,
 401               unsigned long page_size_mask, pgprot_t prot)
 402 {
 403         unsigned long pages = 0;
 404         unsigned long last_map_addr = end;
 405
 406         int i = pmd_index(address);
 407
 408         for (; i < PTRS_PER_PMD; i++, address += PMD_SIZE) {
 409                 unsigned long pte_phys;
 410                 pmd_t *pmd = pmd_page + pmd_index(address);
 411                 pte_t *pte;
 412                 pgprot_t new_prot = prot;
 413
 414                 if (address >= end) {
 415                         if (!after_bootmem) {
 416                                 for (; i < PTRS_PER_PMD; i++, pmd++)
 417                                         set_pmd(pmd, __pmd(0));
 418                         }
 419                         break;
 420                 }
 421
 422                 if (pmd_val(*pmd)) {
 423                         if (!pmd_large(*pmd)) {
 424                                 spin_lock(&init_mm.page_table_lock);
 425                                 last_map_addr = phys_pte_update(pmd, address,
 426                                                                 end, prot);
 427                                 spin_unlock(&init_mm.page_table_lock);
 428                                 continue;
 429                         }
 430                         /*
 431                          * If we are ok with PG_LEVEL_2M mapping, then we will
 432                          * use the existing mapping,
 433                          *
 434                          * Otherwise, we will split the large page mapping but
 435                          * use the same existing protection bits except for
 436                          * large page, so that we don't violate Intel's TLB
 437                          * Application note (317080) which says, while changing
 438                          * the page sizes, new and old translations should
 439                          * not differ with respect to page frame and
 440                          * attributes.
 441                          */
 442                         if (page_size_mask & (1 << PG_LEVEL_2M)) {
 443                                 pages++;
 444                                 continue;
 445                         }
 446                         new_prot = pte_pgprot(pte_clrhuge(*(pte_t *)pmd));
 447                 }
 448
 449                 if (page_size_mask & (1<<PG_LEVEL_2M)) {
 450                         pages++;
 451                         spin_lock(&init_mm.page_table_lock);
 452                         set_pte((pte_t *)pmd,
 453                                 pfn_pte(address >> PAGE_SHIFT,
 454                                         __pgprot(pgprot_val(prot) | _PAGE_PSE)));
 455                         spin_unlock(&init_mm.page_table_lock);
 456                         last_map_addr = (address & PMD_MASK) + PMD_SIZE;
 457                         continue;
 458                 }
 459
 460                 pte = alloc_low_page(&pte_phys);
 461                 last_map_addr = phys_pte_init(pte, address, end, new_prot);
 462                 unmap_low_page(pte);
 463
 464                 spin_lock(&init_mm.page_table_lock);
 465                 pmd_populate_kernel(&init_mm, pmd, __va(pte_phys));
 466                 spin_unlock(&init_mm.page_table_lock);
 467         }
 468         update_page_count(PG_LEVEL_2M, pages);
 469         return last_map_addr;
 470 }
 471
 472 static unsigned long __meminit
 473 phys_pmd_update(pud_t *pud, unsigned long address, unsigned long end,
 474                 unsigned long page_size_mask, pgprot_t prot)
 475 {
 476         pmd_t *pmd = pmd_offset(pud, 0);
 477         unsigned long last_map_addr;
 478
 479         last_map_addr = phys_pmd_init(pmd, address, end, page_size_mask, prot);
 480         __flush_tlb_all();
 481         return last_map_addr;
 482 }
 483
 484 static unsigned long __meminit
 485 phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end,
 486                          unsigned long page_size_mask)
 487 {
 488         unsigned long pages = 0;
 489         unsigned long last_map_addr = end;
 490         int i = pud_index(addr);
 491
 492         for (; i < PTRS_PER_PUD; i++, addr = (addr & PUD_MASK) + PUD_SIZE) {
 493                 unsigned long pmd_phys;
 494                 pud_t *pud = pud_page + pud_index(addr);
 495                 pmd_t *pmd;
 496                 pgprot_t prot = PAGE_KERNEL;
 497
 498                 if (addr >= end)
 499                         break;
 500
 501                 if (!after_bootmem &&
 502                                 !e820_any_mapped(addr, addr+PUD_SIZE, 0)) {
 503                         set_pud(pud, __pud(0));
 504                         continue;
 505                 }
 506
 507                 if (pud_val(*pud)) {
 508                         if (!pud_large(*pud)) {
 509                                 last_map_addr = phys_pmd_update(pud, addr, end,
 510                                                          page_size_mask, prot);
 511                                 continue;
 512                         }
 513                         /*
 514                          * If we are ok with PG_LEVEL_1G mapping, then we will
 515                          * use the existing mapping.
 516                          *
 517                          * Otherwise, we will split the gbpage mapping but use
 518                          * the same existing protection  bits except for large
 519                          * page, so that we don't violate Intel's TLB
 520                          * Application note (317080) which says, while changing
 521                          * the page sizes, new and old translations should
 522                          * not differ with respect to page frame and
 523                          * attributes.
 524                          */
 525                         if (page_size_mask & (1 << PG_LEVEL_1G)) {
 526                                 pages++;
 527                                 continue;
 528                         }
 529                         prot = pte_pgprot(pte_clrhuge(*(pte_t *)pud));
 530                 }
 531
 532                 if (page_size_mask & (1<<PG_LEVEL_1G)) {
 533                         pages++;
 534                         spin_lock(&init_mm.page_table_lock);
 535                         set_pte((pte_t *)pud,
 536                                 pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
 537                         spin_unlock(&init_mm.page_table_lock);
 538                         last_map_addr = (addr & PUD_MASK) + PUD_SIZE;
 539                         continue;
 540                 }
 541
 542                 pmd = alloc_low_page(&pmd_phys);
 543                 last_map_addr = phys_pmd_init(pmd, addr, end, page_size_mask,
 544                                               prot);
 545                 unmap_low_page(pmd);
 546
 547                 spin_lock(&init_mm.page_table_lock);
 548                 pud_populate(&init_mm, pud, __va(pmd_phys));
 549                 spin_unlock(&init_mm.page_table_lock);
 550         }
 551         __flush_tlb_all();
 552
 553         update_page_count(PG_LEVEL_1G, pages);
 554
 555         return last_map_addr;
 556 }
 557
 558 static unsigned long __meminit
 559 phys_pud_update(pgd_t *pgd, unsigned long addr, unsigned long end,
 560                  unsigned long page_size_mask)
 561 {
 562         pud_t *pud;
 563
 564         pud = (pud_t *)pgd_page_vaddr(*pgd);
 565
 566         return phys_pud_init(pud, addr, end, page_size_mask);
 567 }
 568
 569 unsigned long __meminit
 570 kernel_physical_mapping_init(unsigned long start,
 571                              unsigned long end,
 572                              unsigned long page_size_mask)
 573 {
 574         bool pgd_changed = false;
 575         unsigned long next, last_map_addr = end;
 576         unsigned long addr;
 577
 578         start = (unsigned long)__va(start);
 579         end = (unsigned long)__va(end);
 580         addr = start;
 581
 582         for (; start < end; start = next) {
 583                 pgd_t *pgd = pgd_offset_k(start);
 584                 unsigned long pud_phys;
 585                 pud_t *pud;
 586
 587                 next = (start + PGDIR_SIZE) & PGDIR_MASK;
 588                 if (next > end)
 589                         next = end;
 590
 591                 if (pgd_val(*pgd)) {
 592                         last_map_addr = phys_pud_update(pgd, __pa(start),
 593                                                  __pa(end), page_size_mask);
 594                         continue;
 595                 }
 596
 597                 pud = alloc_low_page(&pud_phys);
 598                 last_map_addr = phys_pud_init(pud, __pa(start), __pa(next),
 599                                                  page_size_mask);
 600                 unmap_low_page(pud);
 601
 602                 spin_lock(&init_mm.page_table_lock);
 603                 pgd_populate(&init_mm, pgd, __va(pud_phys));
 604                 spin_unlock(&init_mm.page_table_lock);
 605                 pgd_changed = true;
 606         }
 607
 608         if (pgd_changed)
 609                 sync_global_pgds(addr, end);
 610
 611         __flush_tlb_all();
 612
 613         return last_map_addr;
 614 }
 615
 616 #ifndef CONFIG_NUMA
 617 void __init initmem_init(unsigned long start_pfn, unsigned long end_pfn,
 618                                 int acpi, int k8)
 619 {
 620 #ifndef CONFIG_NO_BOOTMEM
 621         unsigned long bootmap_size, bootmap;
 622
 623         bootmap_size = bootmem_bootmap_pages(end_pfn)<<PAGE_SHIFT;
 624         bootmap = find_e820_area(0, end_pfn<<PAGE_SHIFT, bootmap_size,
 625                                  PAGE_SIZE);
 626         if (bootmap == -1L)
 627                 panic("Cannot find bootmem map of size %ld\n", bootmap_size);
 628         reserve_early(bootmap, bootmap + bootmap_size, "BOOTMAP");
 629         /* don't touch min_low_pfn */
 630         bootmap_size = init_bootmem_node(NODE_DATA(0), bootmap >> PAGE_SHIFT,
 631                                          0, end_pfn);
 632         e820_register_active_regions(0, start_pfn, end_pfn);
 633         free_bootmem_with_active_regions(0, end_pfn);
 634 #else
 635         e820_register_active_regions(0, start_pfn, end_pfn);
 636 #endif
 637 }
 638 #endif
 639
 640 void __init paging_init(void)
 641 {
 642         unsigned long max_zone_pfns[MAX_NR_ZONES];
 643
 644         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 645         max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
 646         max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
 647         max_zone_pfns[ZONE_NORMAL] = max_pfn;
 648
 649         sparse_memory_present_with_active_regions(MAX_NUMNODES);
 650         sparse_init();
 651
 652         /*
 653          * clear the default setting with node 0
 654          * note: don't use nodes_clear here, that is really clearing when
 655          *       numa support is not compiled in, and later node_set_state
 656          *       will not set it back.
 657          */
 658         node_clear_state(0, N_NORMAL_MEMORY);
 659
 660         free_area_init_nodes(max_zone_pfns);
 661 }
 662
 663 /*
 664  * Memory hotplug specific functions
 665  */
 666 #ifdef CONFIG_MEMORY_HOTPLUG
 667 /*
 668  * After memory hotplug the variables max_pfn, max_low_pfn and high_memory need
 669  * updating.
 670  */
 671 static void  update_end_of_memory_vars(u64 start, u64 size)
 672 {
 673         unsigned long end_pfn = PFN_UP(start + size);
 674
 675         if (end_pfn > max_pfn) {
 676                 max_pfn = end_pfn;
 677                 max_low_pfn = end_pfn;
 678                 high_memory = (void *)__va(max_pfn * PAGE_SIZE - 1) + 1;
 679         }
 680 }
 681
 682 /*
 683  * Memory is added always to NORMAL zone. This means you will never get
 684  * additional DMA/DMA32 memory.
 685  */
 686 int arch_add_memory(int nid, u64 start, u64 size)
 687 {
 688         struct pglist_data *pgdat = NODE_DATA(nid);
 689         struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
 690         unsigned long last_mapped_pfn, start_pfn = start >> PAGE_SHIFT;
 691         unsigned long nr_pages = size >> PAGE_SHIFT;
 692         int ret;
 693
 694         last_mapped_pfn = init_memory_mapping(start, start + size);
 695         if (last_mapped_pfn > max_pfn_mapped)
 696                 max_pfn_mapped = last_mapped_pfn;
 697
 698         ret = __add_pages(nid, zone, start_pfn, nr_pages);
 699         WARN_ON_ONCE(ret);
 700
 701         /* update max_pfn, max_low_pfn and high_memory */
 702         update_end_of_memory_vars(start, size);
 703
 704         return ret;
 705 }
 706 EXPORT_SYMBOL_GPL(arch_add_memory);
 707
 708 #if !defined(CONFIG_ACPI_NUMA) && defined(CONFIG_NUMA)
 709 int memory_add_physaddr_to_nid(u64 start)
 710 {
 711         return 0;
 712 }
 713 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
 714 #endif
 715
 716 #endif /* CONFIG_MEMORY_HOTPLUG */
 717
 718 static struct kcore_list kcore_vsyscall;
 719
 720 void __init mem_init(void)
 721 {
 722         long codesize, reservedpages, datasize, initsize;
 723         unsigned long absent_pages;
 724
 725         pci_iommu_alloc();
 726
 727         /* clear_bss() already clear the empty_zero_page */
 728
 729         reservedpages = 0;
 730
 731         /* this will put all low memory onto the freelists */
 732 #ifdef CONFIG_NUMA
 733         totalram_pages = numa_free_all_bootmem();
 734 #else
 735         totalram_pages = free_all_bootmem();
 736 #endif
 737
 738         absent_pages = absent_pages_in_range(0, max_pfn);
 739         reservedpages = max_pfn - totalram_pages - absent_pages;
 740         after_bootmem = 1;
 741
 742         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
 743         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
 744         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
 745
 746         /* Register memory areas for /proc/kcore */
 747         kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
 748                          VSYSCALL_END - VSYSCALL_START, KCORE_OTHER);
 749
 750         printk(KERN_INFO "Memory: %luk/%luk available (%ldk kernel code, "
 751                          "%ldk absent, %ldk reserved, %ldk data, %ldk init)\n",
 752                 nr_free_pages() << (PAGE_SHIFT-10),
 753                 max_pfn << (PAGE_SHIFT-10),
 754                 codesize >> 10,
 755                 absent_pages << (PAGE_SHIFT-10),
 756                 reservedpages << (PAGE_SHIFT-10),
 757                 datasize >> 10,
 758                 initsize >> 10);
 759 }
 760
 761 #ifdef CONFIG_DEBUG_RODATA
 762 const int rodata_test_data = 0xC3;
 763 EXPORT_SYMBOL_GPL(rodata_test_data);
 764
 765 int kernel_set_to_readonly;
 766
 767 void set_kernel_text_rw(void)
 768 {
 769         unsigned long start = PFN_ALIGN(_text);
 770         unsigned long end = PFN_ALIGN(__stop___ex_table);
 771
 772         if (!kernel_set_to_readonly)
 773                 return;
 774
 775         pr_debug("Set kernel text: %lx - %lx for read write\n",
 776                  start, end);
 777
 778         /*
 779          * Make the kernel identity mapping for text RW. Kernel text
 780          * mapping will always be RO. Refer to the comment in
 781          * static_protections() in pageattr.c
 782          */
 783         set_memory_rw(start, (end - start) >> PAGE_SHIFT);
 784 }
 785
 786 void set_kernel_text_ro(void)
 787 {
 788         unsigned long start = PFN_ALIGN(_text);
 789         unsigned long end = PFN_ALIGN(__stop___ex_table);
 790
 791         if (!kernel_set_to_readonly)
 792                 return;
 793
 794         pr_debug("Set kernel text: %lx - %lx for read only\n",
 795                  start, end);
 796
 797         /*
 798          * Set the kernel identity mapping for text RO.
 799          */
 800         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
 801 }
 802
 803 void mark_rodata_ro(void)
 804 {
 805         unsigned long start = PFN_ALIGN(_text);
 806         unsigned long rodata_start =
 807                 ((unsigned long)__start_rodata + PAGE_SIZE - 1) & PAGE_MASK;
 808         unsigned long end = (unsigned long) &__end_rodata_hpage_align;
 809         unsigned long text_end = PAGE_ALIGN((unsigned long) &__stop___ex_table);
 810         unsigned long rodata_end = PAGE_ALIGN((unsigned long) &__end_rodata);
 811         unsigned long data_start = (unsigned long) &_sdata;
 812
 813         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
 814                (end - start) >> 10);
 815         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
 816
 817         kernel_set_to_readonly = 1;
 818
 819         /*
 820          * The rodata section (but not the kernel text!) should also be
 821          * not-executable.
 822          */
 823         set_memory_nx(rodata_start, (end - rodata_start) >> PAGE_SHIFT);
 824
 825         rodata_test();
 826
 827 #ifdef CONFIG_CPA_DEBUG
 828         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
 829         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
 830
 831         printk(KERN_INFO "Testing CPA: again\n");
 832         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
 833 #endif
 834
 835         free_init_pages("unused kernel memory",
 836                         (unsigned long) page_address(virt_to_page(text_end)),
 837                         (unsigned long)
 838                                  page_address(virt_to_page(rodata_start)));
 839         free_init_pages("unused kernel memory",
 840                         (unsigned long) page_address(virt_to_page(rodata_end)),
 841                         (unsigned long) page_address(virt_to_page(data_start)));
 842 }
 843
 844 #endif
 845
 846 int __init reserve_bootmem_generic(unsigned long phys, unsigned long len,
 847                                    int flags)
 848 {
 849 #ifdef CONFIG_NUMA
 850         int nid, next_nid;
 851         int ret;
 852 #endif
 853         unsigned long pfn = phys >> PAGE_SHIFT;
 854
 855         if (pfn >= max_pfn) {
 856                 /*
 857                  * This can happen with kdump kernels when accessing
 858                  * firmware tables:
 859                  */
 860                 if (pfn < max_pfn_mapped)
 861                         return -EFAULT;
 862
 863                 printk(KERN_ERR "reserve_bootmem: illegal reserve %lx %lu\n",
 864                                 phys, len);
 865                 return -EFAULT;
 866         }
 867
 868         /* Should check here against the e820 map to avoid double free */
 869 #ifdef CONFIG_NUMA
 870         nid = phys_to_nid(phys);
 871         next_nid = phys_to_nid(phys + len - 1);
 872         if (nid == next_nid)
 873                 ret = reserve_bootmem_node(NODE_DATA(nid), phys, len, flags);
 874         else
 875                 ret = reserve_bootmem(phys, len, flags);
 876
 877         if (ret != 0)
 878                 return ret;
 879
 880 #else
 881         reserve_bootmem(phys, len, flags);
 882 #endif
 883
 884         if (phys+len <= MAX_DMA_PFN*PAGE_SIZE) {
 885                 dma_reserve += len / PAGE_SIZE;
 886                 set_dma_reserve(dma_reserve);
 887         }
 888
 889         return 0;
 890 }
 891
 892 int kern_addr_valid(unsigned long addr)
 893 {
 894         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
 895         pgd_t *pgd;
 896         pud_t *pud;
 897         pmd_t *pmd;
 898         pte_t *pte;
 899
 900         if (above != 0 && above != -1UL)
 901                 return 0;
 902
 903         pgd = pgd_offset_k(addr);
 904         if (pgd_none(*pgd))
 905                 return 0;
 906
 907         pud = pud_offset(pgd, addr);
 908         if (pud_none(*pud))
 909                 return 0;
 910
 911         pmd = pmd_offset(pud, addr);
 912         if (pmd_none(*pmd))
 913                 return 0;
 914
 915         if (pmd_large(*pmd))
 916                 return pfn_valid(pmd_pfn(*pmd));
 917
 918         pte = pte_offset_kernel(pmd, addr);
 919         if (pte_none(*pte))
 920                 return 0;
 921
 922         return pfn_valid(pte_pfn(*pte));
 923 }
 924
 925 /*
 926  * A pseudo VMA to allow ptrace access for the vsyscall page.  This only
 927  * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
 928  * not need special handling anymore:
 929  */
 930 static struct vm_area_struct gate_vma = {
 931         .vm_start       = VSYSCALL_START,
 932         .vm_end         = VSYSCALL_START + (VSYSCALL_MAPPED_PAGES * PAGE_SIZE),
 933         .vm_page_prot   = PAGE_READONLY_EXEC,
 934         .vm_flags       = VM_READ | VM_EXEC
 935 };
 936
 937 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
 938 {
 939 #ifdef CONFIG_IA32_EMULATION
 940         if (test_tsk_thread_flag(tsk, TIF_IA32))
 941                 return NULL;
 942 #endif
 943         return &gate_vma;
 944 }
 945
 946 int in_gate_area(struct task_struct *task, unsigned long addr)
 947 {
 948         struct vm_area_struct *vma = get_gate_vma(task);
 949
 950         if (!vma)
 951                 return 0;
 952
 953         return (addr >= vma->vm_start) && (addr < vma->vm_end);
 954 }
 955
 956 /*
 957  * Use this when you have no reliable task/vma, typically from interrupt
 958  * context. It is less reliable than using the task's vma and may give
 959  * false positives:
 960  */
 961 int in_gate_area_no_task(unsigned long addr)
 962 {
 963         return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);
 964 }
 965
 966 const char *arch_vma_name(struct vm_area_struct *vma)
 967 {
 968         if (vma->vm_mm && vma->vm_start == (long)vma->vm_mm->context.vdso)
 969                 return "[vdso]";
 970         if (vma == &gate_vma)
 971                 return "[vsyscall]";
 972         return NULL;
 973 }
 974
 975 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 976 /*
 977  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
 978  */
 979 static long __meminitdata addr_start, addr_end;
 980 static void __meminitdata *p_start, *p_end;
 981 static int __meminitdata node_start;
 982
 983 int __meminit
 984 vmemmap_populate(struct page *start_page, unsigned long size, int node)
 985 {
 986         unsigned long addr = (unsigned long)start_page;
 987         unsigned long end = (unsigned long)(start_page + size);
 988         unsigned long next;
 989         pgd_t *pgd;
 990         pud_t *pud;
 991         pmd_t *pmd;
 992
 993         for (; addr < end; addr = next) {
 994                 void *p = NULL;
 995
 996                 pgd = vmemmap_pgd_populate(addr, node);
 997                 if (!pgd)
 998                         return -ENOMEM;
 999
1000                 pud = vmemmap_pud_populate(pgd, addr, node);
1001                 if (!pud)
1002                         return -ENOMEM;
1003
1004                 if (!cpu_has_pse) {
1005                         next = (addr + PAGE_SIZE) & PAGE_MASK;
1006                         pmd = vmemmap_pmd_populate(pud, addr, node);
1007
1008                         if (!pmd)
1009                                 return -ENOMEM;
1010
1011                         p = vmemmap_pte_populate(pmd, addr, node);
1012
1013                         if (!p)
1014                                 return -ENOMEM;
1015
1016                         addr_end = addr + PAGE_SIZE;
1017                         p_end = p + PAGE_SIZE;
1018                 } else {
1019                         next = pmd_addr_end(addr, end);
1020
1021                         pmd = pmd_offset(pud, addr);
1022                         if (pmd_none(*pmd)) {
1023                                 pte_t entry;
1024
1025                                 p = vmemmap_alloc_block_buf(PMD_SIZE, node);
1026                                 if (!p)
1027                                         return -ENOMEM;
1028
1029                                 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
1030                                                 PAGE_KERNEL_LARGE);
1031                                 set_pmd(pmd, __pmd(pte_val(entry)));
1032
1033                                 /* check to see if we have contiguous blocks */
1034                                 if (p_end != p || node_start != node) {
1035                                         if (p_start)
1036                                                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1037                                                        addr_start, addr_end-1, p_start, p_end-1, node_start);
1038                                         addr_start = addr;
1039                                         node_start = node;
1040                                         p_start = p;
1041                                 }
1042
1043                                 addr_end = addr + PMD_SIZE;
1044                                 p_end = p + PMD_SIZE;
1045                         } else
1046                                 vmemmap_verify((pte_t *)pmd, node, addr, next);
1047                 }
1048
1049         }
1050         sync_global_pgds((unsigned long)start_page, end);
1051         return 0;
1052 }
1053
1054 void __meminit vmemmap_populate_print_last(void)
1055 {
1056         if (p_start) {
1057                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
1058                         addr_start, addr_end-1, p_start, p_end-1, node_start);
1059                 p_start = NULL;
1060                 p_end = NULL;
1061                 node_start = 0;
1062         }
1063 }
1064 #endif