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_hotplug.h>
  32 #include <linux/nmi.h>
  33 #include <linux/gfp.h>
  34
  35 #include <asm/processor.h>
  36 #include <asm/bios_ebda.h>
  37 #include <asm/system.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
  56 static int __init parse_direct_gbpages_off(char *arg)
  57 {
  58         direct_gbpages = 0;
  59         return 0;
  60 }
  61 early_param("nogbpages", parse_direct_gbpages_off);
  62
  63 static int __init parse_direct_gbpages_on(char *arg)
  64 {
  65         direct_gbpages = 1;
  66         return 0;
  67 }
  68 early_param("gbpages", parse_direct_gbpages_on);
  69
  70 /*
  71  * NOTE: pagetable_init alloc all the fixmap pagetables contiguous on the
  72  * physical space so we can cache the place of the first one and move
  73  * around without checking the pgd every time.
  74  */
  75
  76 pteval_t __supported_pte_mask __read_mostly = ~_PAGE_IOMAP;
  77 EXPORT_SYMBOL_GPL(__supported_pte_mask);
  78
  79 int force_personality32;
  80
  81 /*
  82  * noexec32=on|off
  83  * Control non executable heap for 32bit processes.
  84  * To control the stack too use noexec=off
  85  *
  86  * on   PROT_READ does not imply PROT_EXEC for 32-bit processes (default)
  87  * off  PROT_READ implies PROT_EXEC
  88  */
  89 static int __init nonx32_setup(char *str)
  90 {
  91         if (!strcmp(str, "on"))
  92                 force_personality32 &= ~READ_IMPLIES_EXEC;
  93         else if (!strcmp(str, "off"))
  94                 force_personality32 |= READ_IMPLIES_EXEC;
  95         return 1;
  96 }
  97 __setup("noexec32=", nonx32_setup);
  98
  99 /*
 100  * When memory was added/removed make sure all the processes MM have
 101  * suitable PGD entries in the local PGD level page.
 102  */
 103 void sync_global_pgds(unsigned long start, unsigned long end)
 104 {
 105         unsigned long address;
 106
 107         for (address = start; address <= end; address += PGDIR_SIZE) {
 108                 const pgd_t *pgd_ref = pgd_offset_k(address);
 109                 struct page *page;
 110
 111                 if (pgd_none(*pgd_ref))
 112                         continue;
 113
 114                 spin_lock(&pgd_lock);
 115                 list_for_each_entry(page, &pgd_list, lru) {
 116                         pgd_t *pgd;
 117                         spinlock_t *pgt_lock;
 118
 119                         pgd = (pgd_t *)page_address(page) + pgd_index(address);
 120                         /* the pgt_lock only for Xen */
 121                         pgt_lock = &pgd_page_get_mm(page)->page_table_lock;
 122                         spin_lock(pgt_lock);
 123
 124                         if (pgd_none(*pgd))
 125                                 set_pgd(pgd, *pgd_ref);
 126                         else
 127                                 BUG_ON(pgd_page_vaddr(*pgd)
 128                                        != pgd_page_vaddr(*pgd_ref));
 129
 130                         spin_unlock(pgt_lock);
 131                 }
 132                 spin_unlock(&pgd_lock);
 133         }
 134 }
 135
 136 /*
 137  * NOTE: This function is marked __ref because it calls __init function
 138  * (alloc_bootmem_pages). It's safe to do it ONLY when after_bootmem == 0.
 139  */
 140 static __ref void *spp_getpage(void)
 141 {
 142         void *ptr;
 143
 144         if (after_bootmem)
 145                 ptr = (void *) get_zeroed_page(GFP_ATOMIC | __GFP_NOTRACK);
 146         else
 147                 ptr = alloc_bootmem_pages(PAGE_SIZE);
 148
 149         if (!ptr || ((unsigned long)ptr & ~PAGE_MASK)) {
 150                 panic("set_pte_phys: cannot allocate page data %s\n",
 151                         after_bootmem ? "after bootmem" : "");
 152         }
 153
 154         pr_debug("spp_getpage %p\n", ptr);
 155
 156         return ptr;
 157 }
 158
 159 static pud_t *fill_pud(pgd_t *pgd, unsigned long vaddr)
 160 {
 161         if (pgd_none(*pgd)) {
 162                 pud_t *pud = (pud_t *)spp_getpage();
 163                 pgd_populate(&init_mm, pgd, pud);
 164                 if (pud != pud_offset(pgd, 0))
 165                         printk(KERN_ERR "PAGETABLE BUG #00! %p <-> %p\n",
 166                                pud, pud_offset(pgd, 0));
 167         }
 168         return pud_offset(pgd, vaddr);
 169 }
 170
 171 static pmd_t *fill_pmd(pud_t *pud, unsigned long vaddr)
 172 {
 173         if (pud_none(*pud)) {
 174                 pmd_t *pmd = (pmd_t *) spp_getpage();
 175                 pud_populate(&init_mm, pud, pmd);
 176                 if (pmd != pmd_offset(pud, 0))
 177                         printk(KERN_ERR "PAGETABLE BUG #01! %p <-> %p\n",
 178                                pmd, pmd_offset(pud, 0));
 179         }
 180         return pmd_offset(pud, vaddr);
 181 }
 182
 183 static pte_t *fill_pte(pmd_t *pmd, unsigned long vaddr)
 184 {
 185         if (pmd_none(*pmd)) {
 186                 pte_t *pte = (pte_t *) spp_getpage();
 187                 pmd_populate_kernel(&init_mm, pmd, pte);
 188                 if (pte != pte_offset_kernel(pmd, 0))
 189                         printk(KERN_ERR "PAGETABLE BUG #02!\n");
 190         }
 191         return pte_offset_kernel(pmd, vaddr);
 192 }
 193
 194 void set_pte_vaddr_pud(pud_t *pud_page, unsigned long vaddr, pte_t new_pte)
 195 {
 196         pud_t *pud;
 197         pmd_t *pmd;
 198         pte_t *pte;
 199
 200         pud = pud_page + pud_index(vaddr);
 201         pmd = fill_pmd(pud, vaddr);
 202         pte = fill_pte(pmd, vaddr);
 203
 204         set_pte(pte, new_pte);
 205
 206         /*
 207          * It's enough to flush this one mapping.
 208          * (PGE mappings get flushed as well)
 209          */
 210         __flush_tlb_one(vaddr);
 211 }
 212
 213 void set_pte_vaddr(unsigned long vaddr, pte_t pteval)
 214 {
 215         pgd_t *pgd;
 216         pud_t *pud_page;
 217
 218         pr_debug("set_pte_vaddr %lx to %lx\n", vaddr, native_pte_val(pteval));
 219
 220         pgd = pgd_offset_k(vaddr);
 221         if (pgd_none(*pgd)) {
 222                 printk(KERN_ERR
 223                         "PGD FIXMAP MISSING, it should be setup in head.S!\n");
 224                 return;
 225         }
 226         pud_page = (pud_t*)pgd_page_vaddr(*pgd);
 227         set_pte_vaddr_pud(pud_page, vaddr, pteval);
 228 }
 229
 230 pmd_t * __init populate_extra_pmd(unsigned long vaddr)
 231 {
 232         pgd_t *pgd;
 233         pud_t *pud;
 234
 235         pgd = pgd_offset_k(vaddr);
 236         pud = fill_pud(pgd, vaddr);
 237         return fill_pmd(pud, vaddr);
 238 }
 239
 240 pte_t * __init populate_extra_pte(unsigned long vaddr)
 241 {
 242         pmd_t *pmd;
 243
 244         pmd = populate_extra_pmd(vaddr);
 245         return fill_pte(pmd, vaddr);
 246 }
 247
 248 /*
 249  * Create large page table mappings for a range of physical addresses.
 250  */
 251 static void __init __init_extra_mapping(unsigned long phys, unsigned long size,
 252                                                 pgprot_t prot)
 253 {
 254         pgd_t *pgd;
 255         pud_t *pud;
 256         pmd_t *pmd;
 257
 258         BUG_ON((phys & ~PMD_MASK) || (size & ~PMD_MASK));
 259         for (; size; phys += PMD_SIZE, size -= PMD_SIZE) {
 260                 pgd = pgd_offset_k((unsigned long)__va(phys));
 261                 if (pgd_none(*pgd)) {
 262                         pud = (pud_t *) spp_getpage();
 263                         set_pgd(pgd, __pgd(__pa(pud) | _KERNPG_TABLE |
 264                                                 _PAGE_USER));
 265                 }
 266                 pud = pud_offset(pgd, (unsigned long)__va(phys));
 267                 if (pud_none(*pud)) {
 268                         pmd = (pmd_t *) spp_getpage();
 269                         set_pud(pud, __pud(__pa(pmd) | _KERNPG_TABLE |
 270                                                 _PAGE_USER));
 271                 }
 272                 pmd = pmd_offset(pud, phys);
 273                 BUG_ON(!pmd_none(*pmd));
 274                 set_pmd(pmd, __pmd(phys | pgprot_val(prot)));
 275         }
 276 }
 277
 278 void __init init_extra_mapping_wb(unsigned long phys, unsigned long size)
 279 {
 280         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE);
 281 }
 282
 283 void __init init_extra_mapping_uc(unsigned long phys, unsigned long size)
 284 {
 285         __init_extra_mapping(phys, size, PAGE_KERNEL_LARGE_NOCACHE);
 286 }
 287
 288 /*
 289  * The head.S code sets up the kernel high mapping:
 290  *
 291  *   from __START_KERNEL_map to __START_KERNEL_map + size (== _end-_text)
 292  *
 293  * phys_addr holds the negative offset to the kernel, which is added
 294  * to the compile time generated pmds. This results in invalid pmds up
 295  * to the point where we hit the physaddr 0 mapping.
 296  *
 297  * We limit the mappings to the region from _text to _end.  _end is
 298  * rounded up to the 2MB boundary. This catches the invalid pmds as
 299  * well, as they are located before _text:
 300  */
 301 void __init cleanup_highmap(void)
 302 {
 303         unsigned long vaddr = __START_KERNEL_map;
 304         unsigned long end = roundup((unsigned long)_end, PMD_SIZE) - 1;
 305         pmd_t *pmd = level2_kernel_pgt;
 306         pmd_t *last_pmd = pmd + PTRS_PER_PMD;
 307
 308         for (; pmd < last_pmd; pmd++, vaddr += PMD_SIZE) {
 309                 if (pmd_none(*pmd))
 310                         continue;
 311                 if (vaddr < (unsigned long) _text || vaddr > end)
 312                         set_pmd(pmd, __pmd(0));
 313         }
 314 }
 315
 316 static __ref void *alloc_low_page(unsigned long *phys)
 317 {
 318         unsigned long pfn = pgt_buf_end++;
 319         void *adr;
 320
 321         if (after_bootmem) {
 322                 adr = (void *)get_zeroed_page(GFP_ATOMIC | __GFP_NOTRACK);
 323                 *phys = __pa(adr);
 324
 325                 return adr;
 326         }
 327
 328         if (pfn >= pgt_buf_top)
 329                 panic("alloc_low_page: ran out of memory");
 330
 331         adr = early_memremap(pfn * PAGE_SIZE, PAGE_SIZE);
 332         clear_page(adr);
 333         *phys  = pfn * PAGE_SIZE;
 334         return adr;
 335 }
 336
 337 static __ref void *map_low_page(void *virt)
 338 {
 339         void *adr;
 340         unsigned long phys, left;
 341
 342         if (after_bootmem)
 343                 return virt;
 344
 345         phys = __pa(virt);
 346         left = phys & (PAGE_SIZE - 1);
 347         adr = early_memremap(phys & PAGE_MASK, PAGE_SIZE);
 348         adr = (void *)(((unsigned long)adr) | left);
 349
 350         return adr;
 351 }
 352
 353 static __ref void unmap_low_page(void *adr)
 354 {
 355         if (after_bootmem)
 356                 return;
 357
 358         early_iounmap((void *)((unsigned long)adr & PAGE_MASK), PAGE_SIZE);
 359 }
 360
 361 static unsigned long __meminit
 362 phys_pte_init(pte_t *pte_page, unsigned long addr, unsigned long end,
 363               pgprot_t prot)
 364 {
 365         unsigned pages = 0;
 366         unsigned long last_map_addr = end;
 367         int i;
 368
 369         pte_t *pte = pte_page + pte_index(addr);
 370
 371         for(i = pte_index(addr); i < PTRS_PER_PTE; i++, addr += PAGE_SIZE, pte++) {
 372
 373                 if (addr >= end) {
 374                         if (!after_bootmem) {
 375                                 for(; i < PTRS_PER_PTE; i++, pte++)
 376                                         set_pte(pte, __pte(0));
 377                         }
 378                         break;
 379                 }
 380
 381                 /*
 382                  * We will re-use the existing mapping.
 383                  * Xen for example has some special requirements, like mapping
 384                  * pagetable pages as RO. So assume someone who pre-setup
 385                  * these mappings are more intelligent.
 386                  */
 387                 if (pte_val(*pte)) {
 388                         pages++;
 389                         continue;
 390                 }
 391
 392                 if (0)
 393                         printk("   pte=%p addr=%lx pte=%016lx\n",
 394                                pte, addr, pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL).pte);
 395                 pages++;
 396                 set_pte(pte, pfn_pte(addr >> PAGE_SHIFT, prot));
 397                 last_map_addr = (addr & PAGE_MASK) + PAGE_SIZE;
 398         }
 399
 400         update_page_count(PG_LEVEL_4K, pages);
 401
 402         return last_map_addr;
 403 }
 404
 405 static unsigned long __meminit
 406 phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end,
 407               unsigned long page_size_mask, pgprot_t prot)
 408 {
 409         unsigned long pages = 0;
 410         unsigned long last_map_addr = end;
 411
 412         int i = pmd_index(address);
 413
 414         for (; i < PTRS_PER_PMD; i++, address += PMD_SIZE) {
 415                 unsigned long pte_phys;
 416                 pmd_t *pmd = pmd_page + pmd_index(address);
 417                 pte_t *pte;
 418                 pgprot_t new_prot = prot;
 419
 420                 if (address >= end) {
 421                         if (!after_bootmem) {
 422                                 for (; i < PTRS_PER_PMD; i++, pmd++)
 423                                         set_pmd(pmd, __pmd(0));
 424                         }
 425                         break;
 426                 }
 427
 428                 if (pmd_val(*pmd)) {
 429                         if (!pmd_large(*pmd)) {
 430                                 spin_lock(&init_mm.page_table_lock);
 431                                 pte = map_low_page((pte_t *)pmd_page_vaddr(*pmd));
 432                                 last_map_addr = phys_pte_init(pte, address,
 433                                                                 end, prot);
 434                                 unmap_low_page(pte);
 435                                 spin_unlock(&init_mm.page_table_lock);
 436                                 continue;
 437                         }
 438                         /*
 439                          * If we are ok with PG_LEVEL_2M mapping, then we will
 440                          * use the existing mapping,
 441                          *
 442                          * Otherwise, we will split the large page mapping but
 443                          * use the same existing protection bits except for
 444                          * large page, so that we don't violate Intel's TLB
 445                          * Application note (317080) which says, while changing
 446                          * the page sizes, new and old translations should
 447                          * not differ with respect to page frame and
 448                          * attributes.
 449                          */
 450                         if (page_size_mask & (1 << PG_LEVEL_2M)) {
 451                                 pages++;
 452                                 continue;
 453                         }
 454                         new_prot = pte_pgprot(pte_clrhuge(*(pte_t *)pmd));
 455                 }
 456
 457                 if (page_size_mask & (1<<PG_LEVEL_2M)) {
 458                         pages++;
 459                         spin_lock(&init_mm.page_table_lock);
 460                         set_pte((pte_t *)pmd,
 461                                 pfn_pte(address >> PAGE_SHIFT,
 462                                         __pgprot(pgprot_val(prot) | _PAGE_PSE)));
 463                         spin_unlock(&init_mm.page_table_lock);
 464                         last_map_addr = (address & PMD_MASK) + PMD_SIZE;
 465                         continue;
 466                 }
 467
 468                 pte = alloc_low_page(&pte_phys);
 469                 last_map_addr = phys_pte_init(pte, address, end, new_prot);
 470                 unmap_low_page(pte);
 471
 472                 spin_lock(&init_mm.page_table_lock);
 473                 pmd_populate_kernel(&init_mm, pmd, __va(pte_phys));
 474                 spin_unlock(&init_mm.page_table_lock);
 475         }
 476         update_page_count(PG_LEVEL_2M, pages);
 477         return last_map_addr;
 478 }
 479
 480 static unsigned long __meminit
 481 phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end,
 482                          unsigned long page_size_mask)
 483 {
 484         unsigned long pages = 0;
 485         unsigned long last_map_addr = end;
 486         int i = pud_index(addr);
 487
 488         for (; i < PTRS_PER_PUD; i++, addr = (addr & PUD_MASK) + PUD_SIZE) {
 489                 unsigned long pmd_phys;
 490                 pud_t *pud = pud_page + pud_index(addr);
 491                 pmd_t *pmd;
 492                 pgprot_t prot = PAGE_KERNEL;
 493
 494                 if (addr >= end)
 495                         break;
 496
 497                 if (!after_bootmem &&
 498                                 !e820_any_mapped(addr, addr+PUD_SIZE, 0)) {
 499                         set_pud(pud, __pud(0));
 500                         continue;
 501                 }
 502
 503                 if (pud_val(*pud)) {
 504                         if (!pud_large(*pud)) {
 505                                 pmd = map_low_page(pmd_offset(pud, 0));
 506                                 last_map_addr = phys_pmd_init(pmd, addr, end,
 507                                                          page_size_mask, prot);
 508                                 unmap_low_page(pmd);
 509                                 __flush_tlb_all();
 510                                 continue;
 511                         }
 512                         /*
 513                          * If we are ok with PG_LEVEL_1G mapping, then we will
 514                          * use the existing mapping.
 515                          *
 516                          * Otherwise, we will split the gbpage mapping but use
 517                          * the same existing protection  bits except for large
 518                          * page, so that we don't violate Intel's TLB
 519                          * Application note (317080) which says, while changing
 520                          * the page sizes, new and old translations should
 521                          * not differ with respect to page frame and
 522                          * attributes.
 523                          */
 524                         if (page_size_mask & (1 << PG_LEVEL_1G)) {
 525                                 pages++;
 526                                 continue;
 527                         }
 528                         prot = pte_pgprot(pte_clrhuge(*(pte_t *)pud));
 529                 }
 530
 531                 if (page_size_mask & (1<<PG_LEVEL_1G)) {
 532                         pages++;
 533                         spin_lock(&init_mm.page_table_lock);
 534                         set_pte((pte_t *)pud,
 535                                 pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL_LARGE));
 536                         spin_unlock(&init_mm.page_table_lock);
 537                         last_map_addr = (addr & PUD_MASK) + PUD_SIZE;
 538                         continue;
 539                 }
 540
 541                 pmd = alloc_low_page(&pmd_phys);
 542                 last_map_addr = phys_pmd_init(pmd, addr, end, page_size_mask,
 543                                               prot);
 544                 unmap_low_page(pmd);
 545
 546                 spin_lock(&init_mm.page_table_lock);
 547                 pud_populate(&init_mm, pud, __va(pmd_phys));
 548                 spin_unlock(&init_mm.page_table_lock);
 549         }
 550         __flush_tlb_all();
 551
 552         update_page_count(PG_LEVEL_1G, pages);
 553
 554         return last_map_addr;
 555 }
 556
 557 unsigned long __meminit
 558 kernel_physical_mapping_init(unsigned long start,
 559                              unsigned long end,
 560                              unsigned long page_size_mask)
 561 {
 562         bool pgd_changed = false;
 563         unsigned long next, last_map_addr = end;
 564         unsigned long addr;
 565
 566         start = (unsigned long)__va(start);
 567         end = (unsigned long)__va(end);
 568         addr = start;
 569
 570         for (; start < end; start = next) {
 571                 pgd_t *pgd = pgd_offset_k(start);
 572                 unsigned long pud_phys;
 573                 pud_t *pud;
 574
 575                 next = (start + PGDIR_SIZE) & PGDIR_MASK;
 576                 if (next > end)
 577                         next = end;
 578
 579                 if (pgd_val(*pgd)) {
 580                         pud = map_low_page((pud_t *)pgd_page_vaddr(*pgd));
 581                         last_map_addr = phys_pud_init(pud, __pa(start),
 582                                                  __pa(end), page_size_mask);
 583                         unmap_low_page(pud);
 584                         continue;
 585                 }
 586
 587                 pud = alloc_low_page(&pud_phys);
 588                 last_map_addr = phys_pud_init(pud, __pa(start), __pa(next),
 589                                                  page_size_mask);
 590                 unmap_low_page(pud);
 591
 592                 spin_lock(&init_mm.page_table_lock);
 593                 pgd_populate(&init_mm, pgd, __va(pud_phys));
 594                 spin_unlock(&init_mm.page_table_lock);
 595                 pgd_changed = true;
 596         }
 597
 598         if (pgd_changed)
 599                 sync_global_pgds(addr, end);
 600
 601         __flush_tlb_all();
 602
 603         return last_map_addr;
 604 }
 605
 606 #ifndef CONFIG_NUMA
 607 void __init initmem_init(void)
 608 {
 609         memblock_x86_register_active_regions(0, 0, max_pfn);
 610 }
 611 #endif
 612
 613 void __init paging_init(void)
 614 {
 615         unsigned long max_zone_pfns[MAX_NR_ZONES];
 616
 617         memset(max_zone_pfns, 0, sizeof(max_zone_pfns));
 618         max_zone_pfns[ZONE_DMA] = MAX_DMA_PFN;
 619         max_zone_pfns[ZONE_DMA32] = MAX_DMA32_PFN;
 620         max_zone_pfns[ZONE_NORMAL] = max_pfn;
 621
 622         sparse_memory_present_with_active_regions(MAX_NUMNODES);
 623         sparse_init();
 624
 625         /*
 626          * clear the default setting with node 0
 627          * note: don't use nodes_clear here, that is really clearing when
 628          *       numa support is not compiled in, and later node_set_state
 629          *       will not set it back.
 630          */
 631         node_clear_state(0, N_NORMAL_MEMORY);
 632
 633         free_area_init_nodes(max_zone_pfns);
 634 }
 635
 636 /*
 637  * Memory hotplug specific functions
 638  */
 639 #ifdef CONFIG_MEMORY_HOTPLUG
 640 /*
 641  * After memory hotplug the variables max_pfn, max_low_pfn and high_memory need
 642  * updating.
 643  */
 644 static void  update_end_of_memory_vars(u64 start, u64 size)
 645 {
 646         unsigned long end_pfn = PFN_UP(start + size);
 647
 648         if (end_pfn > max_pfn) {
 649                 max_pfn = end_pfn;
 650                 max_low_pfn = end_pfn;
 651                 high_memory = (void *)__va(max_pfn * PAGE_SIZE - 1) + 1;
 652         }
 653 }
 654
 655 /*
 656  * Memory is added always to NORMAL zone. This means you will never get
 657  * additional DMA/DMA32 memory.
 658  */
 659 int arch_add_memory(int nid, u64 start, u64 size)
 660 {
 661         struct pglist_data *pgdat = NODE_DATA(nid);
 662         struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
 663         unsigned long last_mapped_pfn, start_pfn = start >> PAGE_SHIFT;
 664         unsigned long nr_pages = size >> PAGE_SHIFT;
 665         int ret;
 666
 667         last_mapped_pfn = init_memory_mapping(start, start + size);
 668         if (last_mapped_pfn > max_pfn_mapped)
 669                 max_pfn_mapped = last_mapped_pfn;
 670
 671         ret = __add_pages(nid, zone, start_pfn, nr_pages);
 672         WARN_ON_ONCE(ret);
 673
 674         /* update max_pfn, max_low_pfn and high_memory */
 675         update_end_of_memory_vars(start, size);
 676
 677         return ret;
 678 }
 679 EXPORT_SYMBOL_GPL(arch_add_memory);
 680
 681 #if !defined(CONFIG_ACPI_NUMA) && defined(CONFIG_NUMA)
 682 int memory_add_physaddr_to_nid(u64 start)
 683 {
 684         return 0;
 685 }
 686 EXPORT_SYMBOL_GPL(memory_add_physaddr_to_nid);
 687 #endif
 688
 689 #endif /* CONFIG_MEMORY_HOTPLUG */
 690
 691 static struct kcore_list kcore_vsyscall;
 692
 693 void __init mem_init(void)
 694 {
 695         long codesize, reservedpages, datasize, initsize;
 696         unsigned long absent_pages;
 697
 698         pci_iommu_alloc();
 699
 700         /* clear_bss() already clear the empty_zero_page */
 701
 702         reservedpages = 0;
 703
 704         /* this will put all low memory onto the freelists */
 705 #ifdef CONFIG_NUMA
 706         totalram_pages = numa_free_all_bootmem();
 707 #else
 708         totalram_pages = free_all_bootmem();
 709 #endif
 710
 711         absent_pages = absent_pages_in_range(0, max_pfn);
 712         reservedpages = max_pfn - totalram_pages - absent_pages;
 713         after_bootmem = 1;
 714
 715         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
 716         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
 717         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
 718
 719         /* Register memory areas for /proc/kcore */
 720         kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
 721                          VSYSCALL_END - VSYSCALL_START, KCORE_OTHER);
 722
 723         printk(KERN_INFO "Memory: %luk/%luk available (%ldk kernel code, "
 724                          "%ldk absent, %ldk reserved, %ldk data, %ldk init)\n",
 725                 nr_free_pages() << (PAGE_SHIFT-10),
 726                 max_pfn << (PAGE_SHIFT-10),
 727                 codesize >> 10,
 728                 absent_pages << (PAGE_SHIFT-10),
 729                 reservedpages << (PAGE_SHIFT-10),
 730                 datasize >> 10,
 731                 initsize >> 10);
 732 }
 733
 734 #ifdef CONFIG_DEBUG_RODATA
 735 const int rodata_test_data = 0xC3;
 736 EXPORT_SYMBOL_GPL(rodata_test_data);
 737
 738 int kernel_set_to_readonly;
 739
 740 void set_kernel_text_rw(void)
 741 {
 742         unsigned long start = PFN_ALIGN(_text);
 743         unsigned long end = PFN_ALIGN(__stop___ex_table);
 744
 745         if (!kernel_set_to_readonly)
 746                 return;
 747
 748         pr_debug("Set kernel text: %lx - %lx for read write\n",
 749                  start, end);
 750
 751         /*
 752          * Make the kernel identity mapping for text RW. Kernel text
 753          * mapping will always be RO. Refer to the comment in
 754          * static_protections() in pageattr.c
 755          */
 756         set_memory_rw(start, (end - start) >> PAGE_SHIFT);
 757 }
 758
 759 void set_kernel_text_ro(void)
 760 {
 761         unsigned long start = PFN_ALIGN(_text);
 762         unsigned long end = PFN_ALIGN(__stop___ex_table);
 763
 764         if (!kernel_set_to_readonly)
 765                 return;
 766
 767         pr_debug("Set kernel text: %lx - %lx for read only\n",
 768                  start, end);
 769
 770         /*
 771          * Set the kernel identity mapping for text RO.
 772          */
 773         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
 774 }
 775
 776 void mark_rodata_ro(void)
 777 {
 778         unsigned long start = PFN_ALIGN(_text);
 779         unsigned long rodata_start =
 780                 ((unsigned long)__start_rodata + PAGE_SIZE - 1) & PAGE_MASK;
 781         unsigned long end = (unsigned long) &__end_rodata_hpage_align;
 782         unsigned long text_end = PAGE_ALIGN((unsigned long) &__stop___ex_table);
 783         unsigned long rodata_end = PAGE_ALIGN((unsigned long) &__end_rodata);
 784         unsigned long data_start = (unsigned long) &_sdata;
 785
 786         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
 787                (end - start) >> 10);
 788         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
 789
 790         kernel_set_to_readonly = 1;
 791
 792         /*
 793          * The rodata section (but not the kernel text!) should also be
 794          * not-executable.
 795          */
 796         set_memory_nx(rodata_start, (end - rodata_start) >> PAGE_SHIFT);
 797
 798         rodata_test();
 799
 800 #ifdef CONFIG_CPA_DEBUG
 801         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
 802         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
 803
 804         printk(KERN_INFO "Testing CPA: again\n");
 805         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
 806 #endif
 807
 808         free_init_pages("unused kernel memory",
 809                         (unsigned long) page_address(virt_to_page(text_end)),
 810                         (unsigned long)
 811                                  page_address(virt_to_page(rodata_start)));
 812         free_init_pages("unused kernel memory",
 813                         (unsigned long) page_address(virt_to_page(rodata_end)),
 814                         (unsigned long) page_address(virt_to_page(data_start)));
 815 }
 816
 817 #endif
 818
 819 int kern_addr_valid(unsigned long addr)
 820 {
 821         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
 822         pgd_t *pgd;
 823         pud_t *pud;
 824         pmd_t *pmd;
 825         pte_t *pte;
 826
 827         if (above != 0 && above != -1UL)
 828                 return 0;
 829
 830         pgd = pgd_offset_k(addr);
 831         if (pgd_none(*pgd))
 832                 return 0;
 833
 834         pud = pud_offset(pgd, addr);
 835         if (pud_none(*pud))
 836                 return 0;
 837
 838         pmd = pmd_offset(pud, addr);
 839         if (pmd_none(*pmd))
 840                 return 0;
 841
 842         if (pmd_large(*pmd))
 843                 return pfn_valid(pmd_pfn(*pmd));
 844
 845         pte = pte_offset_kernel(pmd, addr);
 846         if (pte_none(*pte))
 847                 return 0;
 848
 849         return pfn_valid(pte_pfn(*pte));
 850 }
 851
 852 /*
 853  * A pseudo VMA to allow ptrace access for the vsyscall page.  This only
 854  * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
 855  * not need special handling anymore:
 856  */
 857 static struct vm_area_struct gate_vma = {
 858         .vm_start       = VSYSCALL_START,
 859         .vm_end         = VSYSCALL_START + (VSYSCALL_MAPPED_PAGES * PAGE_SIZE),
 860         .vm_page_prot   = PAGE_READONLY_EXEC,
 861         .vm_flags       = VM_READ | VM_EXEC
 862 };
 863
 864 struct vm_area_struct *get_gate_vma(struct task_struct *tsk)
 865 {
 866 #ifdef CONFIG_IA32_EMULATION
 867         if (test_tsk_thread_flag(tsk, TIF_IA32))
 868                 return NULL;
 869 #endif
 870         return &gate_vma;
 871 }
 872
 873 int in_gate_area(struct task_struct *task, unsigned long addr)
 874 {
 875         struct vm_area_struct *vma = get_gate_vma(task);
 876
 877         if (!vma)
 878                 return 0;
 879
 880         return (addr >= vma->vm_start) && (addr < vma->vm_end);
 881 }
 882
 883 /*
 884  * Use this when you have no reliable task/vma, typically from interrupt
 885  * context. It is less reliable than using the task's vma and may give
 886  * false positives:
 887  */
 888 int in_gate_area_no_task(unsigned long addr)
 889 {
 890         return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);
 891 }
 892
 893 const char *arch_vma_name(struct vm_area_struct *vma)
 894 {
 895         if (vma->vm_mm && vma->vm_start == (long)vma->vm_mm->context.vdso)
 896                 return "[vdso]";
 897         if (vma == &gate_vma)
 898                 return "[vsyscall]";
 899         return NULL;
 900 }
 901
 902 #ifdef CONFIG_X86_UV
 903 #define MIN_MEMORY_BLOCK_SIZE   (1 << SECTION_SIZE_BITS)
 904
 905 unsigned long memory_block_size_bytes(void)
 906 {
 907         if (is_uv_system()) {
 908                 printk(KERN_INFO "UV: memory block size 2GB\n");
 909                 return 2UL * 1024 * 1024 * 1024;
 910         }
 911         return MIN_MEMORY_BLOCK_SIZE;
 912 }
 913 #endif
 914
 915 #ifdef CONFIG_SPARSEMEM_VMEMMAP
 916 /*
 917  * Initialise the sparsemem vmemmap using huge-pages at the PMD level.
 918  */
 919 static long __meminitdata addr_start, addr_end;
 920 static void __meminitdata *p_start, *p_end;
 921 static int __meminitdata node_start;
 922
 923 int __meminit
 924 vmemmap_populate(struct page *start_page, unsigned long size, int node)
 925 {
 926         unsigned long addr = (unsigned long)start_page;
 927         unsigned long end = (unsigned long)(start_page + size);
 928         unsigned long next;
 929         pgd_t *pgd;
 930         pud_t *pud;
 931         pmd_t *pmd;
 932
 933         for (; addr < end; addr = next) {
 934                 void *p = NULL;
 935
 936                 pgd = vmemmap_pgd_populate(addr, node);
 937                 if (!pgd)
 938                         return -ENOMEM;
 939
 940                 pud = vmemmap_pud_populate(pgd, addr, node);
 941                 if (!pud)
 942                         return -ENOMEM;
 943
 944                 if (!cpu_has_pse) {
 945                         next = (addr + PAGE_SIZE) & PAGE_MASK;
 946                         pmd = vmemmap_pmd_populate(pud, addr, node);
 947
 948                         if (!pmd)
 949                                 return -ENOMEM;
 950
 951                         p = vmemmap_pte_populate(pmd, addr, node);
 952
 953                         if (!p)
 954                                 return -ENOMEM;
 955
 956                         addr_end = addr + PAGE_SIZE;
 957                         p_end = p + PAGE_SIZE;
 958                 } else {
 959                         next = pmd_addr_end(addr, end);
 960
 961                         pmd = pmd_offset(pud, addr);
 962                         if (pmd_none(*pmd)) {
 963                                 pte_t entry;
 964
 965                                 p = vmemmap_alloc_block_buf(PMD_SIZE, node);
 966                                 if (!p)
 967                                         return -ENOMEM;
 968
 969                                 entry = pfn_pte(__pa(p) >> PAGE_SHIFT,
 970                                                 PAGE_KERNEL_LARGE);
 971                                 set_pmd(pmd, __pmd(pte_val(entry)));
 972
 973                                 /* check to see if we have contiguous blocks */
 974                                 if (p_end != p || node_start != node) {
 975                                         if (p_start)
 976                                                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
 977                                                        addr_start, addr_end-1, p_start, p_end-1, node_start);
 978                                         addr_start = addr;
 979                                         node_start = node;
 980                                         p_start = p;
 981                                 }
 982
 983                                 addr_end = addr + PMD_SIZE;
 984                                 p_end = p + PMD_SIZE;
 985                         } else
 986                                 vmemmap_verify((pte_t *)pmd, node, addr, next);
 987                 }
 988
 989         }
 990         sync_global_pgds((unsigned long)start_page, end);
 991         return 0;
 992 }
 993
 994 void __meminit vmemmap_populate_print_last(void)
 995 {
 996         if (p_start) {
 997                 printk(KERN_DEBUG " [%lx-%lx] PMD -> [%p-%p] on node %d\n",
 998                         addr_start, addr_end-1, p_start, p_end-1, node_start);
 999                 p_start = NULL;
1000                 p_end = NULL;
1001                 node_start = 0;
1002         }
1003 }
1004 #endif