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 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                 struct page *page;
 111
 112                 if (pgd_none(*pgd_ref))
 113                         continue;
 114
 115                 spin_lock(&pgd_lock);
 116                 list_for_each_entry(page, &pgd_list, lru) {
 117                         pgd_t *pgd;
 118                         spinlock_t *pgt_lock;
 119
 120                         pgd = (pgd_t *)page_address(page) + pgd_index(address);
 121                         /* the pgt_lock only for Xen */
 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(&pgd_lock);
 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 _brk_end.  _brk_end
 299  * is 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 vaddr_end = __START_KERNEL_map + (max_pfn_mapped << PAGE_SHIFT);
 306         unsigned long end = roundup((unsigned long)_brk_end, PMD_SIZE) - 1;
 307         pmd_t *pmd = level2_kernel_pgt;
 308
 309         for (; vaddr + PMD_SIZE - 1 < vaddr_end; 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;
 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 ((pgt_buf_end + 1) >= pgt_buf_top) {
 330                 unsigned long ret;
 331                 if (min_pfn_mapped >= max_pfn_mapped)
 332                         panic("alloc_low_page: ran out of memory");
 333                 ret = memblock_find_in_range(min_pfn_mapped << PAGE_SHIFT,
 334                                         max_pfn_mapped << PAGE_SHIFT,
 335                                         PAGE_SIZE, PAGE_SIZE);
 336                 if (!ret)
 337                         panic("alloc_low_page: can not alloc memory");
 338                 memblock_reserve(ret, PAGE_SIZE);
 339                 pfn = ret >> PAGE_SHIFT;
 340         } else
 341                 pfn = pgt_buf_end++;
 342
 343         adr = early_memremap(pfn * PAGE_SIZE, PAGE_SIZE);
 344         clear_page(adr);
 345         *phys  = pfn * PAGE_SIZE;
 346         return adr;
 347 }
 348
 349 static __ref void *map_low_page(void *virt)
 350 {
 351         void *adr;
 352         unsigned long phys, left;
 353
 354         if (after_bootmem)
 355                 return virt;
 356
 357         phys = __pa(virt);
 358         left = phys & (PAGE_SIZE - 1);
 359         adr = early_memremap(phys & PAGE_MASK, PAGE_SIZE);
 360         adr = (void *)(((unsigned long)adr) | left);
 361
 362         return adr;
 363 }
 364
 365 static __ref void unmap_low_page(void *adr)
 366 {
 367         if (after_bootmem)
 368                 return;
 369
 370         early_iounmap((void *)((unsigned long)adr & PAGE_MASK), PAGE_SIZE);
 371 }
 372
 373 static unsigned long __meminit
 374 phys_pte_init(pte_t *pte_page, unsigned long addr, unsigned long end,
 375               pgprot_t prot)
 376 {
 377         unsigned long pages = 0, next;
 378         unsigned long last_map_addr = end;
 379         int i;
 380
 381         pte_t *pte = pte_page + pte_index(addr);
 382
 383         for (i = pte_index(addr); i < PTRS_PER_PTE; i++, addr = next, pte++) {
 384                 next = (addr & PAGE_MASK) + PAGE_SIZE;
 385                 if (addr >= end) {
 386                         if (!after_bootmem &&
 387                             !e820_any_mapped(addr & PAGE_MASK, next, E820_RAM) &&
 388                             !e820_any_mapped(addr & PAGE_MASK, next, E820_RESERVED_KERN))
 389                                 set_pte(pte, __pte(0));
 390                         continue;
 391                 }
 392
 393                 /*
 394                  * We will re-use the existing mapping.
 395                  * Xen for example has some special requirements, like mapping
 396                  * pagetable pages as RO. So assume someone who pre-setup
 397                  * these mappings are more intelligent.
 398                  */
 399                 if (pte_val(*pte)) {
 400                         if (!after_bootmem)
 401                                 pages++;
 402                         continue;
 403                 }
 404
 405                 if (0)
 406                         printk("   pte=%p addr=%lx pte=%016lx\n",
 407                                pte, addr, pfn_pte(addr >> PAGE_SHIFT, PAGE_KERNEL).pte);
 408                 pages++;
 409                 set_pte(pte, pfn_pte(addr >> PAGE_SHIFT, prot));
 410                 last_map_addr = (addr & PAGE_MASK) + PAGE_SIZE;
 411         }
 412
 413         update_page_count(PG_LEVEL_4K, pages);
 414
 415         return last_map_addr;
 416 }
 417
 418 static unsigned long __meminit
 419 phys_pmd_init(pmd_t *pmd_page, unsigned long address, unsigned long end,
 420               unsigned long page_size_mask, pgprot_t prot)
 421 {
 422         unsigned long pages = 0, next;
 423         unsigned long last_map_addr = end;
 424
 425         int i = pmd_index(address);
 426
 427         for (; i < PTRS_PER_PMD; i++, address = next) {
 428                 unsigned long pte_phys;
 429                 pmd_t *pmd = pmd_page + pmd_index(address);
 430                 pte_t *pte;
 431                 pgprot_t new_prot = prot;
 432
 433                 next = (address & PMD_MASK) + PMD_SIZE;
 434                 if (address >= end) {
 435                         if (!after_bootmem &&
 436                             !e820_any_mapped(address & PMD_MASK, next, E820_RAM) &&
 437                             !e820_any_mapped(address & PMD_MASK, next, E820_RESERVED_KERN))
 438                                 set_pmd(pmd, __pmd(0));
 439                         continue;
 440                 }
 441
 442                 if (pmd_val(*pmd)) {
 443                         if (!pmd_large(*pmd)) {
 444                                 spin_lock(&init_mm.page_table_lock);
 445                                 pte = map_low_page((pte_t *)pmd_page_vaddr(*pmd));
 446                                 last_map_addr = phys_pte_init(pte, address,
 447                                                                 end, prot);
 448                                 unmap_low_page(pte);
 449                                 spin_unlock(&init_mm.page_table_lock);
 450                                 continue;
 451                         }
 452                         /*
 453                          * If we are ok with PG_LEVEL_2M mapping, then we will
 454                          * use the existing mapping,
 455                          *
 456                          * Otherwise, we will split the large page mapping but
 457                          * use the same existing protection bits except for
 458                          * large page, so that we don't violate Intel's TLB
 459                          * Application note (317080) which says, while changing
 460                          * the page sizes, new and old translations should
 461                          * not differ with respect to page frame and
 462                          * attributes.
 463                          */
 464                         if (page_size_mask & (1 << PG_LEVEL_2M)) {
 465                                 if (!after_bootmem)
 466                                         pages++;
 467                                 last_map_addr = next;
 468                                 continue;
 469                         }
 470                         new_prot = pte_pgprot(pte_clrhuge(*(pte_t *)pmd));
 471                 }
 472
 473                 if (page_size_mask & (1<<PG_LEVEL_2M)) {
 474                         pages++;
 475                         spin_lock(&init_mm.page_table_lock);
 476                         set_pte((pte_t *)pmd,
 477                                 pfn_pte((address & PMD_MASK) >> PAGE_SHIFT,
 478                                         __pgprot(pgprot_val(prot) | _PAGE_PSE)));
 479                         spin_unlock(&init_mm.page_table_lock);
 480                         last_map_addr = next;
 481                         continue;
 482                 }
 483
 484                 pte = alloc_low_page(&pte_phys);
 485                 last_map_addr = phys_pte_init(pte, address, end, new_prot);
 486                 unmap_low_page(pte);
 487
 488                 spin_lock(&init_mm.page_table_lock);
 489                 pmd_populate_kernel(&init_mm, pmd, __va(pte_phys));
 490                 spin_unlock(&init_mm.page_table_lock);
 491         }
 492         update_page_count(PG_LEVEL_2M, pages);
 493         return last_map_addr;
 494 }
 495
 496 static unsigned long __meminit
 497 phys_pud_init(pud_t *pud_page, unsigned long addr, unsigned long end,
 498                          unsigned long page_size_mask)
 499 {
 500         unsigned long pages = 0, next;
 501         unsigned long last_map_addr = end;
 502         int i = pud_index(addr);
 503
 504         for (; i < PTRS_PER_PUD; i++, addr = next) {
 505                 unsigned long pmd_phys;
 506                 pud_t *pud = pud_page + pud_index(addr);
 507                 pmd_t *pmd;
 508                 pgprot_t prot = PAGE_KERNEL;
 509
 510                 next = (addr & PUD_MASK) + PUD_SIZE;
 511                 if (addr >= end) {
 512                         if (!after_bootmem &&
 513                             !e820_any_mapped(addr & PUD_MASK, next, E820_RAM) &&
 514                             !e820_any_mapped(addr & PUD_MASK, next, E820_RESERVED_KERN))
 515                                 set_pud(pud, __pud(0));
 516                         continue;
 517                 }
 518
 519                 if (pud_val(*pud)) {
 520                         if (!pud_large(*pud)) {
 521                                 pmd = map_low_page(pmd_offset(pud, 0));
 522                                 last_map_addr = phys_pmd_init(pmd, addr, end,
 523                                                          page_size_mask, prot);
 524                                 unmap_low_page(pmd);
 525                                 __flush_tlb_all();
 526                                 continue;
 527                         }
 528                         /*
 529                          * If we are ok with PG_LEVEL_1G mapping, then we will
 530                          * use the existing mapping.
 531                          *
 532                          * Otherwise, we will split the gbpage mapping but use
 533                          * the same existing protection  bits except for large
 534                          * page, so that we don't violate Intel's TLB
 535                          * Application note (317080) which says, while changing
 536                          * the page sizes, new and old translations should
 537                          * not differ with respect to page frame and
 538                          * attributes.
 539                          */
 540                         if (page_size_mask & (1 << PG_LEVEL_1G)) {
 541                                 if (!after_bootmem)
 542                                         pages++;
 543                                 last_map_addr = next;
 544                                 continue;
 545                         }
 546                         prot = pte_pgprot(pte_clrhuge(*(pte_t *)pud));
 547                 }
 548
 549                 if (page_size_mask & (1<<PG_LEVEL_1G)) {
 550                         pages++;
 551                         spin_lock(&init_mm.page_table_lock);
 552                         set_pte((pte_t *)pud,
 553                                 pfn_pte((addr & PUD_MASK) >> PAGE_SHIFT,
 554                                         PAGE_KERNEL_LARGE));
 555                         spin_unlock(&init_mm.page_table_lock);
 556                         last_map_addr = next;
 557                         continue;
 558                 }
 559
 560                 pmd = alloc_low_page(&pmd_phys);
 561                 last_map_addr = phys_pmd_init(pmd, addr, end, page_size_mask,
 562                                               prot);
 563                 unmap_low_page(pmd);
 564
 565                 spin_lock(&init_mm.page_table_lock);
 566                 pud_populate(&init_mm, pud, __va(pmd_phys));
 567                 spin_unlock(&init_mm.page_table_lock);
 568         }
 569         __flush_tlb_all();
 570
 571         update_page_count(PG_LEVEL_1G, pages);
 572
 573         return last_map_addr;
 574 }
 575
 576 unsigned long __meminit
 577 kernel_physical_mapping_init(unsigned long start,
 578                              unsigned long end,
 579                              unsigned long page_size_mask)
 580 {
 581         bool pgd_changed = false;
 582         unsigned long next, last_map_addr = end;
 583         unsigned long addr;
 584
 585         start = (unsigned long)__va(start);
 586         end = (unsigned long)__va(end);
 587         addr = start;
 588
 589         for (; start < end; start = next) {
 590                 pgd_t *pgd = pgd_offset_k(start);
 591                 unsigned long pud_phys;
 592                 pud_t *pud;
 593
 594                 next = (start + PGDIR_SIZE) & PGDIR_MASK;
 595                 if (next > end)
 596                         next = end;
 597
 598                 if (pgd_val(*pgd)) {
 599                         pud = map_low_page((pud_t *)pgd_page_vaddr(*pgd));
 600                         last_map_addr = phys_pud_init(pud, __pa(start),
 601                                                  __pa(end), page_size_mask);
 602                         unmap_low_page(pud);
 603                         continue;
 604                 }
 605
 606                 pud = alloc_low_page(&pud_phys);
 607                 last_map_addr = phys_pud_init(pud, __pa(start), __pa(next),
 608                                                  page_size_mask);
 609                 unmap_low_page(pud);
 610
 611                 spin_lock(&init_mm.page_table_lock);
 612                 pgd_populate(&init_mm, pgd, __va(pud_phys));
 613                 spin_unlock(&init_mm.page_table_lock);
 614                 pgd_changed = true;
 615         }
 616
 617         if (pgd_changed)
 618                 sync_global_pgds(addr, end);
 619
 620         __flush_tlb_all();
 621
 622         return last_map_addr;
 623 }
 624
 625 #ifndef CONFIG_NUMA
 626 void __init initmem_init(void)
 627 {
 628         memblock_set_node(0, (phys_addr_t)ULLONG_MAX, 0);
 629 }
 630 #endif
 631
 632 void __init paging_init(void)
 633 {
 634         sparse_memory_present_with_active_regions(MAX_NUMNODES);
 635         sparse_init();
 636
 637         /*
 638          * clear the default setting with node 0
 639          * note: don't use nodes_clear here, that is really clearing when
 640          *       numa support is not compiled in, and later node_set_state
 641          *       will not set it back.
 642          */
 643         node_clear_state(0, N_NORMAL_MEMORY);
 644
 645         zone_sizes_init();
 646 }
 647
 648 /*
 649  * Memory hotplug specific functions
 650  */
 651 #ifdef CONFIG_MEMORY_HOTPLUG
 652 /*
 653  * After memory hotplug the variables max_pfn, max_low_pfn and high_memory need
 654  * updating.
 655  */
 656 static void  update_end_of_memory_vars(u64 start, u64 size)
 657 {
 658         unsigned long end_pfn = PFN_UP(start + size);
 659
 660         if (end_pfn > max_pfn) {
 661                 max_pfn = end_pfn;
 662                 max_low_pfn = end_pfn;
 663                 high_memory = (void *)__va(max_pfn * PAGE_SIZE - 1) + 1;
 664         }
 665 }
 666
 667 /*
 668  * Memory is added always to NORMAL zone. This means you will never get
 669  * additional DMA/DMA32 memory.
 670  */
 671 int arch_add_memory(int nid, u64 start, u64 size)
 672 {
 673         struct pglist_data *pgdat = NODE_DATA(nid);
 674         struct zone *zone = pgdat->node_zones + ZONE_NORMAL;
 675         unsigned long start_pfn = start >> PAGE_SHIFT;
 676         unsigned long nr_pages = size >> PAGE_SHIFT;
 677         int ret;
 678
 679         init_memory_mapping(start, start + size);
 680
 681         ret = __add_pages(nid, zone, start_pfn, nr_pages);
 682         WARN_ON_ONCE(ret);
 683
 684         /* update max_pfn, max_low_pfn and high_memory */
 685         update_end_of_memory_vars(start, size);
 686
 687         return ret;
 688 }
 689 EXPORT_SYMBOL_GPL(arch_add_memory);
 690
 691 #endif /* CONFIG_MEMORY_HOTPLUG */
 692
 693 static struct kcore_list kcore_vsyscall;
 694
 695 void __init mem_init(void)
 696 {
 697         long codesize, reservedpages, datasize, initsize;
 698         unsigned long absent_pages;
 699
 700         pci_iommu_alloc();
 701
 702         /* clear_bss() already clear the empty_zero_page */
 703
 704         reservedpages = 0;
 705
 706         /* this will put all low memory onto the freelists */
 707 #ifdef CONFIG_NUMA
 708         totalram_pages = numa_free_all_bootmem();
 709 #else
 710         totalram_pages = free_all_bootmem();
 711 #endif
 712
 713         absent_pages = absent_pages_in_range(0, max_pfn);
 714         reservedpages = max_pfn - totalram_pages - absent_pages;
 715         after_bootmem = 1;
 716
 717         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
 718         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
 719         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
 720
 721         /* Register memory areas for /proc/kcore */
 722         kclist_add(&kcore_vsyscall, (void *)VSYSCALL_START,
 723                          VSYSCALL_END - VSYSCALL_START, KCORE_OTHER);
 724
 725         printk(KERN_INFO "Memory: %luk/%luk available (%ldk kernel code, "
 726                          "%ldk absent, %ldk reserved, %ldk data, %ldk init)\n",
 727                 nr_free_pages() << (PAGE_SHIFT-10),
 728                 max_pfn << (PAGE_SHIFT-10),
 729                 codesize >> 10,
 730                 absent_pages << (PAGE_SHIFT-10),
 731                 reservedpages << (PAGE_SHIFT-10),
 732                 datasize >> 10,
 733                 initsize >> 10);
 734 }
 735
 736 #ifdef CONFIG_DEBUG_RODATA
 737 const int rodata_test_data = 0xC3;
 738 EXPORT_SYMBOL_GPL(rodata_test_data);
 739
 740 int kernel_set_to_readonly;
 741
 742 void set_kernel_text_rw(void)
 743 {
 744         unsigned long start = PFN_ALIGN(_text);
 745         unsigned long end = PFN_ALIGN(__stop___ex_table);
 746
 747         if (!kernel_set_to_readonly)
 748                 return;
 749
 750         pr_debug("Set kernel text: %lx - %lx for read write\n",
 751                  start, end);
 752
 753         /*
 754          * Make the kernel identity mapping for text RW. Kernel text
 755          * mapping will always be RO. Refer to the comment in
 756          * static_protections() in pageattr.c
 757          */
 758         set_memory_rw(start, (end - start) >> PAGE_SHIFT);
 759 }
 760
 761 void set_kernel_text_ro(void)
 762 {
 763         unsigned long start = PFN_ALIGN(_text);
 764         unsigned long end = PFN_ALIGN(__stop___ex_table);
 765
 766         if (!kernel_set_to_readonly)
 767                 return;
 768
 769         pr_debug("Set kernel text: %lx - %lx for read only\n",
 770                  start, end);
 771
 772         /*
 773          * Set the kernel identity mapping for text RO.
 774          */
 775         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
 776 }
 777
 778 void mark_rodata_ro(void)
 779 {
 780         unsigned long start = PFN_ALIGN(_text);
 781         unsigned long rodata_start =
 782                 ((unsigned long)__start_rodata + PAGE_SIZE - 1) & PAGE_MASK;
 783         unsigned long end = (unsigned long) &__end_rodata_hpage_align;
 784         unsigned long text_end = PAGE_ALIGN((unsigned long) &__stop___ex_table);
 785         unsigned long rodata_end = PAGE_ALIGN((unsigned long) &__end_rodata);
 786         unsigned long data_start = (unsigned long) &_sdata;
 787
 788         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
 789                (end - start) >> 10);
 790         set_memory_ro(start, (end - start) >> PAGE_SHIFT);
 791
 792         kernel_set_to_readonly = 1;
 793
 794         /*
 795          * The rodata section (but not the kernel text!) should also be
 796          * not-executable.
 797          */
 798         set_memory_nx(rodata_start, (end - rodata_start) >> PAGE_SHIFT);
 799
 800         rodata_test();
 801
 802 #ifdef CONFIG_CPA_DEBUG
 803         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, end);
 804         set_memory_rw(start, (end-start) >> PAGE_SHIFT);
 805
 806         printk(KERN_INFO "Testing CPA: again\n");
 807         set_memory_ro(start, (end-start) >> PAGE_SHIFT);
 808 #endif
 809
 810         free_init_pages("unused kernel memory",
 811                         (unsigned long) page_address(virt_to_page(text_end)),
 812                         (unsigned long)
 813                                  page_address(virt_to_page(rodata_start)));
 814         free_init_pages("unused kernel memory",
 815                         (unsigned long) page_address(virt_to_page(rodata_end)),
 816                         (unsigned long) page_address(virt_to_page(data_start)));
 817 }
 818
 819 #endif
 820
 821 int kern_addr_valid(unsigned long addr)
 822 {
 823         unsigned long above = ((long)addr) >> __VIRTUAL_MASK_SHIFT;
 824         pgd_t *pgd;
 825         pud_t *pud;
 826         pmd_t *pmd;
 827         pte_t *pte;
 828
 829         if (above != 0 && above != -1UL)
 830                 return 0;
 831
 832         pgd = pgd_offset_k(addr);
 833         if (pgd_none(*pgd))
 834                 return 0;
 835
 836         pud = pud_offset(pgd, addr);
 837         if (pud_none(*pud))
 838                 return 0;
 839
 840         pmd = pmd_offset(pud, addr);
 841         if (pmd_none(*pmd))
 842                 return 0;
 843
 844         if (pmd_large(*pmd))
 845                 return pfn_valid(pmd_pfn(*pmd));
 846
 847         pte = pte_offset_kernel(pmd, addr);
 848         if (pte_none(*pte))
 849                 return 0;
 850
 851         return pfn_valid(pte_pfn(*pte));
 852 }
 853
 854 /*
 855  * A pseudo VMA to allow ptrace access for the vsyscall page.  This only
 856  * covers the 64bit vsyscall page now. 32bit has a real VMA now and does
 857  * not need special handling anymore:
 858  */
 859 static struct vm_area_struct gate_vma = {
 860         .vm_start       = VSYSCALL_START,
 861         .vm_end         = VSYSCALL_START + (VSYSCALL_MAPPED_PAGES * PAGE_SIZE),
 862         .vm_page_prot   = PAGE_READONLY_EXEC,
 863         .vm_flags       = VM_READ | VM_EXEC
 864 };
 865
 866 struct vm_area_struct *get_gate_vma(struct mm_struct *mm)
 867 {
 868 #ifdef CONFIG_IA32_EMULATION
 869         if (!mm || mm->context.ia32_compat)
 870                 return NULL;
 871 #endif
 872         return &gate_vma;
 873 }
 874
 875 int in_gate_area(struct mm_struct *mm, unsigned long addr)
 876 {
 877         struct vm_area_struct *vma = get_gate_vma(mm);
 878
 879         if (!vma)
 880                 return 0;
 881
 882         return (addr >= vma->vm_start) && (addr < vma->vm_end);
 883 }
 884
 885 /*
 886  * Use this when you have no reliable mm, typically from interrupt
 887  * context. It is less reliable than using a task's mm and may give
 888  * false positives.
 889  */
 890 int in_gate_area_no_mm(unsigned long addr)
 891 {
 892         return (addr >= VSYSCALL_START) && (addr < VSYSCALL_END);
 893 }
 894
 895 const char *arch_vma_name(struct vm_area_struct *vma)
 896 {
 897         if (vma->vm_mm && vma->vm_start == (long)vma->vm_mm->context.vdso)
 898                 return "[vdso]";
 899         if (vma == &gate_vma)
 900                 return "[vsyscall]";
 901         return NULL;
 902 }
 903
 904 #ifdef CONFIG_X86_UV
 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