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