arch/x86/mm/init_32.c

   1 /*
   2  *
   3  *  Copyright (C) 1995  Linus Torvalds
   4  *
   5  *  Support of BIGMEM added by Gerhard Wichert, Siemens AG, July 1999
   6  */
   7
   8 #include <linux/module.h>
   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/hugetlb.h>
  19 #include <linux/swap.h>
  20 #include <linux/smp.h>
  21 #include <linux/init.h>
  22 #include <linux/highmem.h>
  23 #include <linux/pagemap.h>
  24 #include <linux/pci.h>
  25 #include <linux/pfn.h>
  26 #include <linux/poison.h>
  27 #include <linux/bootmem.h>
  28 #include <linux/memblock.h>
  29 #include <linux/proc_fs.h>
  30 #include <linux/memory_hotplug.h>
  31 #include <linux/initrd.h>
  32 #include <linux/cpumask.h>
  33 #include <linux/gfp.h>
  34
  35 #include <asm/asm.h>
  36 #include <asm/bios_ebda.h>
  37 #include <asm/processor.h>
  38 #include <asm/uaccess.h>
  39 #include <asm/pgtable.h>
  40 #include <asm/dma.h>
  41 #include <asm/fixmap.h>
  42 #include <asm/e820.h>
  43 #include <asm/apic.h>
  44 #include <asm/bugs.h>
  45 #include <asm/tlb.h>
  46 #include <asm/tlbflush.h>
  47 #include <asm/olpc_ofw.h>
  48 #include <asm/pgalloc.h>
  49 #include <asm/sections.h>
  50 #include <asm/paravirt.h>
  51 #include <asm/setup.h>
  52 #include <asm/cacheflush.h>
  53 #include <asm/page_types.h>
  54 #include <asm/init.h>
  55
  56 #include "mm_internal.h"
  57
  58 unsigned long highstart_pfn, highend_pfn;
  59
  60 static noinline int do_test_wp_bit(void);
  61
  62 bool __read_mostly __vmalloc_start_set = false;
  63
  64 /*
  65  * Creates a middle page table and puts a pointer to it in the
  66  * given global directory entry. This only returns the gd entry
  67  * in non-PAE compilation mode, since the middle layer is folded.
  68  */
  69 static pmd_t * __init one_md_table_init(pgd_t *pgd)
  70 {
  71         pud_t *pud;
  72         pmd_t *pmd_table;
  73
  74 #ifdef CONFIG_X86_PAE
  75         if (!(pgd_val(*pgd) & _PAGE_PRESENT)) {
  76                 pmd_table = (pmd_t *)alloc_low_page();
  77                 paravirt_alloc_pmd(&init_mm, __pa(pmd_table) >> PAGE_SHIFT);
  78                 set_pgd(pgd, __pgd(__pa(pmd_table) | _PAGE_PRESENT));
  79                 pud = pud_offset(pgd, 0);
  80                 BUG_ON(pmd_table != pmd_offset(pud, 0));
  81
  82                 return pmd_table;
  83         }
  84 #endif
  85         pud = pud_offset(pgd, 0);
  86         pmd_table = pmd_offset(pud, 0);
  87
  88         return pmd_table;
  89 }
  90
  91 /*
  92  * Create a page table and place a pointer to it in a middle page
  93  * directory entry:
  94  */
  95 static pte_t * __init one_page_table_init(pmd_t *pmd)
  96 {
  97         if (!(pmd_val(*pmd) & _PAGE_PRESENT)) {
  98                 pte_t *page_table = (pte_t *)alloc_low_page();
  99
 100                 paravirt_alloc_pte(&init_mm, __pa(page_table) >> PAGE_SHIFT);
 101                 set_pmd(pmd, __pmd(__pa(page_table) | _PAGE_TABLE));
 102                 BUG_ON(page_table != pte_offset_kernel(pmd, 0));
 103         }
 104
 105         return pte_offset_kernel(pmd, 0);
 106 }
 107
 108 pmd_t * __init populate_extra_pmd(unsigned long vaddr)
 109 {
 110         int pgd_idx = pgd_index(vaddr);
 111         int pmd_idx = pmd_index(vaddr);
 112
 113         return one_md_table_init(swapper_pg_dir + pgd_idx) + pmd_idx;
 114 }
 115
 116 pte_t * __init populate_extra_pte(unsigned long vaddr)
 117 {
 118         int pte_idx = pte_index(vaddr);
 119         pmd_t *pmd;
 120
 121         pmd = populate_extra_pmd(vaddr);
 122         return one_page_table_init(pmd) + pte_idx;
 123 }
 124
 125 static unsigned long __init
 126 page_table_range_init_count(unsigned long start, unsigned long end)
 127 {
 128         unsigned long count = 0;
 129 #ifdef CONFIG_HIGHMEM
 130         int pmd_idx_kmap_begin = fix_to_virt(FIX_KMAP_END) >> PMD_SHIFT;
 131         int pmd_idx_kmap_end = fix_to_virt(FIX_KMAP_BEGIN) >> PMD_SHIFT;
 132         int pgd_idx, pmd_idx;
 133         unsigned long vaddr;
 134
 135         if (pmd_idx_kmap_begin == pmd_idx_kmap_end)
 136                 return 0;
 137
 138         vaddr = start;
 139         pgd_idx = pgd_index(vaddr);
 140
 141         for ( ; (pgd_idx < PTRS_PER_PGD) && (vaddr != end); pgd_idx++) {
 142                 for (; (pmd_idx < PTRS_PER_PMD) && (vaddr != end);
 143                                                         pmd_idx++) {
 144                         if ((vaddr >> PMD_SHIFT) >= pmd_idx_kmap_begin &&
 145                             (vaddr >> PMD_SHIFT) <= pmd_idx_kmap_end)
 146                                 count++;
 147                         vaddr += PMD_SIZE;
 148                 }
 149                 pmd_idx = 0;
 150         }
 151 #endif
 152         return count;
 153 }
 154
 155 static pte_t *__init page_table_kmap_check(pte_t *pte, pmd_t *pmd,
 156                                            unsigned long vaddr, pte_t *lastpte,
 157                                            void **adr)
 158 {
 159 #ifdef CONFIG_HIGHMEM
 160         /*
 161          * Something (early fixmap) may already have put a pte
 162          * page here, which causes the page table allocation
 163          * to become nonlinear. Attempt to fix it, and if it
 164          * is still nonlinear then we have to bug.
 165          */
 166         int pmd_idx_kmap_begin = fix_to_virt(FIX_KMAP_END) >> PMD_SHIFT;
 167         int pmd_idx_kmap_end = fix_to_virt(FIX_KMAP_BEGIN) >> PMD_SHIFT;
 168
 169         if (pmd_idx_kmap_begin != pmd_idx_kmap_end
 170             && (vaddr >> PMD_SHIFT) >= pmd_idx_kmap_begin
 171             && (vaddr >> PMD_SHIFT) <= pmd_idx_kmap_end) {
 172                 pte_t *newpte;
 173                 int i;
 174
 175                 BUG_ON(after_bootmem);
 176                 newpte = *adr;
 177                 for (i = 0; i < PTRS_PER_PTE; i++)
 178                         set_pte(newpte + i, pte[i]);
 179                 *adr = (void *)(((unsigned long)(*adr)) + PAGE_SIZE);
 180
 181                 paravirt_alloc_pte(&init_mm, __pa(newpte) >> PAGE_SHIFT);
 182                 set_pmd(pmd, __pmd(__pa(newpte)|_PAGE_TABLE));
 183                 BUG_ON(newpte != pte_offset_kernel(pmd, 0));
 184                 __flush_tlb_all();
 185
 186                 paravirt_release_pte(__pa(pte) >> PAGE_SHIFT);
 187                 pte = newpte;
 188         }
 189         BUG_ON(vaddr < fix_to_virt(FIX_KMAP_BEGIN - 1)
 190                && vaddr > fix_to_virt(FIX_KMAP_END)
 191                && lastpte && lastpte + PTRS_PER_PTE != pte);
 192 #endif
 193         return pte;
 194 }
 195
 196 /*
 197  * This function initializes a certain range of kernel virtual memory
 198  * with new bootmem page tables, everywhere page tables are missing in
 199  * the given range.
 200  *
 201  * NOTE: The pagetables are allocated contiguous on the physical space
 202  * so we can cache the place of the first one and move around without
 203  * checking the pgd every time.
 204  */
 205 static void __init
 206 page_table_range_init(unsigned long start, unsigned long end, pgd_t *pgd_base)
 207 {
 208         int pgd_idx, pmd_idx;
 209         unsigned long vaddr;
 210         pgd_t *pgd;
 211         pmd_t *pmd;
 212         pte_t *pte = NULL;
 213         unsigned long count = page_table_range_init_count(start, end);
 214         void *adr = NULL;
 215
 216         if (count)
 217                 adr = alloc_low_pages(count);
 218
 219         vaddr = start;
 220         pgd_idx = pgd_index(vaddr);
 221         pmd_idx = pmd_index(vaddr);
 222         pgd = pgd_base + pgd_idx;
 223
 224         for ( ; (pgd_idx < PTRS_PER_PGD) && (vaddr != end); pgd++, pgd_idx++) {
 225                 pmd = one_md_table_init(pgd);
 226                 pmd = pmd + pmd_index(vaddr);
 227                 for (; (pmd_idx < PTRS_PER_PMD) && (vaddr != end);
 228                                                         pmd++, pmd_idx++) {
 229                         pte = page_table_kmap_check(one_page_table_init(pmd),
 230                                                     pmd, vaddr, pte, &adr);
 231
 232                         vaddr += PMD_SIZE;
 233                 }
 234                 pmd_idx = 0;
 235         }
 236 }
 237
 238 static inline int is_kernel_text(unsigned long addr)
 239 {
 240         if (addr >= (unsigned long)_text && addr <= (unsigned long)__init_end)
 241                 return 1;
 242         return 0;
 243 }
 244
 245 /*
 246  * This maps the physical memory to kernel virtual address space, a total
 247  * of max_low_pfn pages, by creating page tables starting from address
 248  * PAGE_OFFSET:
 249  */
 250 unsigned long __init
 251 kernel_physical_mapping_init(unsigned long start,
 252                              unsigned long end,
 253                              unsigned long page_size_mask)
 254 {
 255         int use_pse = page_size_mask == (1<<PG_LEVEL_2M);
 256         unsigned long last_map_addr = end;
 257         unsigned long start_pfn, end_pfn;
 258         pgd_t *pgd_base = swapper_pg_dir;
 259         int pgd_idx, pmd_idx, pte_ofs;
 260         unsigned long pfn;
 261         pgd_t *pgd;
 262         pmd_t *pmd;
 263         pte_t *pte;
 264         unsigned pages_2m, pages_4k;
 265         int mapping_iter;
 266
 267         start_pfn = start >> PAGE_SHIFT;
 268         end_pfn = end >> PAGE_SHIFT;
 269
 270         /*
 271          * First iteration will setup identity mapping using large/small pages
 272          * based on use_pse, with other attributes same as set by
 273          * the early code in head_32.S
 274          *
 275          * Second iteration will setup the appropriate attributes (NX, GLOBAL..)
 276          * as desired for the kernel identity mapping.
 277          *
 278          * This two pass mechanism conforms to the TLB app note which says:
 279          *
 280          *     "Software should not write to a paging-structure entry in a way
 281          *      that would change, for any linear address, both the page size
 282          *      and either the page frame or attributes."
 283          */
 284         mapping_iter = 1;
 285
 286         if (!cpu_has_pse)
 287                 use_pse = 0;
 288
 289 repeat:
 290         pages_2m = pages_4k = 0;
 291         pfn = start_pfn;
 292         pgd_idx = pgd_index((pfn<<PAGE_SHIFT) + PAGE_OFFSET);
 293         pgd = pgd_base + pgd_idx;
 294         for (; pgd_idx < PTRS_PER_PGD; pgd++, pgd_idx++) {
 295                 pmd = one_md_table_init(pgd);
 296
 297                 if (pfn >= end_pfn)
 298                         continue;
 299 #ifdef CONFIG_X86_PAE
 300                 pmd_idx = pmd_index((pfn<<PAGE_SHIFT) + PAGE_OFFSET);
 301                 pmd += pmd_idx;
 302 #else
 303                 pmd_idx = 0;
 304 #endif
 305                 for (; pmd_idx < PTRS_PER_PMD && pfn < end_pfn;
 306                      pmd++, pmd_idx++) {
 307                         unsigned int addr = pfn * PAGE_SIZE + PAGE_OFFSET;
 308
 309                         /*
 310                          * Map with big pages if possible, otherwise
 311                          * create normal page tables:
 312                          */
 313                         if (use_pse) {
 314                                 unsigned int addr2;
 315                                 pgprot_t prot = PAGE_KERNEL_LARGE;
 316                                 /*
 317                                  * first pass will use the same initial
 318                                  * identity mapping attribute + _PAGE_PSE.
 319                                  */
 320                                 pgprot_t init_prot =
 321                                         __pgprot(PTE_IDENT_ATTR |
 322                                                  _PAGE_PSE);
 323
 324                                 pfn &= PMD_MASK >> PAGE_SHIFT;
 325                                 addr2 = (pfn + PTRS_PER_PTE-1) * PAGE_SIZE +
 326                                         PAGE_OFFSET + PAGE_SIZE-1;
 327
 328                                 if (is_kernel_text(addr) ||
 329                                     is_kernel_text(addr2))
 330                                         prot = PAGE_KERNEL_LARGE_EXEC;
 331
 332                                 pages_2m++;
 333                                 if (mapping_iter == 1)
 334                                         set_pmd(pmd, pfn_pmd(pfn, init_prot));
 335                                 else
 336                                         set_pmd(pmd, pfn_pmd(pfn, prot));
 337
 338                                 pfn += PTRS_PER_PTE;
 339                                 continue;
 340                         }
 341                         pte = one_page_table_init(pmd);
 342
 343                         pte_ofs = pte_index((pfn<<PAGE_SHIFT) + PAGE_OFFSET);
 344                         pte += pte_ofs;
 345                         for (; pte_ofs < PTRS_PER_PTE && pfn < end_pfn;
 346                              pte++, pfn++, pte_ofs++, addr += PAGE_SIZE) {
 347                                 pgprot_t prot = PAGE_KERNEL;
 348                                 /*
 349                                  * first pass will use the same initial
 350                                  * identity mapping attribute.
 351                                  */
 352                                 pgprot_t init_prot = __pgprot(PTE_IDENT_ATTR);
 353
 354                                 if (is_kernel_text(addr))
 355                                         prot = PAGE_KERNEL_EXEC;
 356
 357                                 pages_4k++;
 358                                 if (mapping_iter == 1) {
 359                                         set_pte(pte, pfn_pte(pfn, init_prot));
 360                                         last_map_addr = (pfn << PAGE_SHIFT) + PAGE_SIZE;
 361                                 } else
 362                                         set_pte(pte, pfn_pte(pfn, prot));
 363                         }
 364                 }
 365         }
 366         if (mapping_iter == 1) {
 367                 /*
 368                  * update direct mapping page count only in the first
 369                  * iteration.
 370                  */
 371                 update_page_count(PG_LEVEL_2M, pages_2m);
 372                 update_page_count(PG_LEVEL_4K, pages_4k);
 373
 374                 /*
 375                  * local global flush tlb, which will flush the previous
 376                  * mappings present in both small and large page TLB's.
 377                  */
 378                 __flush_tlb_all();
 379
 380                 /*
 381                  * Second iteration will set the actual desired PTE attributes.
 382                  */
 383                 mapping_iter = 2;
 384                 goto repeat;
 385         }
 386         return last_map_addr;
 387 }
 388
 389 pte_t *kmap_pte;
 390 pgprot_t kmap_prot;
 391
 392 static inline pte_t *kmap_get_fixmap_pte(unsigned long vaddr)
 393 {
 394         return pte_offset_kernel(pmd_offset(pud_offset(pgd_offset_k(vaddr),
 395                         vaddr), vaddr), vaddr);
 396 }
 397
 398 static void __init kmap_init(void)
 399 {
 400         unsigned long kmap_vstart;
 401
 402         /*
 403          * Cache the first kmap pte:
 404          */
 405         kmap_vstart = __fix_to_virt(FIX_KMAP_BEGIN);
 406         kmap_pte = kmap_get_fixmap_pte(kmap_vstart);
 407
 408         kmap_prot = PAGE_KERNEL;
 409 }
 410
 411 #ifdef CONFIG_HIGHMEM
 412 static void __init permanent_kmaps_init(pgd_t *pgd_base)
 413 {
 414         unsigned long vaddr;
 415         pgd_t *pgd;
 416         pud_t *pud;
 417         pmd_t *pmd;
 418         pte_t *pte;
 419
 420         vaddr = PKMAP_BASE;
 421         page_table_range_init(vaddr, vaddr + PAGE_SIZE*LAST_PKMAP, pgd_base);
 422
 423         pgd = swapper_pg_dir + pgd_index(vaddr);
 424         pud = pud_offset(pgd, vaddr);
 425         pmd = pmd_offset(pud, vaddr);
 426         pte = pte_offset_kernel(pmd, vaddr);
 427         pkmap_page_table = pte;
 428 }
 429
 430 static void __init add_one_highpage_init(struct page *page)
 431 {
 432         ClearPageReserved(page);
 433         init_page_count(page);
 434         __free_page(page);
 435         totalhigh_pages++;
 436 }
 437
 438 void __init add_highpages_with_active_regions(int nid,
 439                          unsigned long start_pfn, unsigned long end_pfn)
 440 {
 441         phys_addr_t start, end;
 442         u64 i;
 443
 444         for_each_free_mem_range(i, nid, &start, &end, NULL) {
 445                 unsigned long pfn = clamp_t(unsigned long, PFN_UP(start),
 446                                             start_pfn, end_pfn);
 447                 unsigned long e_pfn = clamp_t(unsigned long, PFN_DOWN(end),
 448                                               start_pfn, end_pfn);
 449                 for ( ; pfn < e_pfn; pfn++)
 450                         if (pfn_valid(pfn))
 451                                 add_one_highpage_init(pfn_to_page(pfn));
 452         }
 453 }
 454 #else
 455 static inline void permanent_kmaps_init(pgd_t *pgd_base)
 456 {
 457 }
 458 #endif /* CONFIG_HIGHMEM */
 459
 460 void __init native_pagetable_init(void)
 461 {
 462         unsigned long pfn, va;
 463         pgd_t *pgd, *base = swapper_pg_dir;
 464         pud_t *pud;
 465         pmd_t *pmd;
 466         pte_t *pte;
 467
 468         /*
 469          * Remove any mappings which extend past the end of physical
 470          * memory from the boot time page table.
 471          * In virtual address space, we should have at least two pages
 472          * from VMALLOC_END to pkmap or fixmap according to VMALLOC_END
 473          * definition. And max_low_pfn is set to VMALLOC_END physical
 474          * address. If initial memory mapping is doing right job, we
 475          * should have pte used near max_low_pfn or one pmd is not present.
 476          */
 477         for (pfn = max_low_pfn; pfn < 1<<(32-PAGE_SHIFT); pfn++) {
 478                 va = PAGE_OFFSET + (pfn<<PAGE_SHIFT);
 479                 pgd = base + pgd_index(va);
 480                 if (!pgd_present(*pgd))
 481                         break;
 482
 483                 pud = pud_offset(pgd, va);
 484                 pmd = pmd_offset(pud, va);
 485                 if (!pmd_present(*pmd))
 486                         break;
 487
 488                 /* should not be large page here */
 489                 if (pmd_large(*pmd)) {
 490                         pr_warn("try to clear pte for ram above max_low_pfn: pfn: %lx pmd: %p pmd phys: %lx, but pmd is big page and is not using pte !\n",
 491                                 pfn, pmd, __pa(pmd));
 492                         BUG_ON(1);
 493                 }
 494
 495                 pte = pte_offset_kernel(pmd, va);
 496                 if (!pte_present(*pte))
 497                         break;
 498
 499                 printk(KERN_DEBUG "clearing pte for ram above max_low_pfn: pfn: %lx pmd: %p pmd phys: %lx pte: %p pte phys: %lx\n",
 500                                 pfn, pmd, __pa(pmd), pte, __pa(pte));
 501                 pte_clear(NULL, va, pte);
 502         }
 503         paravirt_alloc_pmd(&init_mm, __pa(base) >> PAGE_SHIFT);
 504         paging_init();
 505 }
 506
 507 /*
 508  * Build a proper pagetable for the kernel mappings.  Up until this
 509  * point, we've been running on some set of pagetables constructed by
 510  * the boot process.
 511  *
 512  * If we're booting on native hardware, this will be a pagetable
 513  * constructed in arch/x86/kernel/head_32.S.  The root of the
 514  * pagetable will be swapper_pg_dir.
 515  *
 516  * If we're booting paravirtualized under a hypervisor, then there are
 517  * more options: we may already be running PAE, and the pagetable may
 518  * or may not be based in swapper_pg_dir.  In any case,
 519  * paravirt_pagetable_init() will set up swapper_pg_dir
 520  * appropriately for the rest of the initialization to work.
 521  *
 522  * In general, pagetable_init() assumes that the pagetable may already
 523  * be partially populated, and so it avoids stomping on any existing
 524  * mappings.
 525  */
 526 void __init early_ioremap_page_table_range_init(void)
 527 {
 528         pgd_t *pgd_base = swapper_pg_dir;
 529         unsigned long vaddr, end;
 530
 531         /*
 532          * Fixed mappings, only the page table structure has to be
 533          * created - mappings will be set by set_fixmap():
 534          */
 535         vaddr = __fix_to_virt(__end_of_fixed_addresses - 1) & PMD_MASK;
 536         end = (FIXADDR_TOP + PMD_SIZE - 1) & PMD_MASK;
 537         page_table_range_init(vaddr, end, pgd_base);
 538         early_ioremap_reset();
 539 }
 540
 541 static void __init pagetable_init(void)
 542 {
 543         pgd_t *pgd_base = swapper_pg_dir;
 544
 545         permanent_kmaps_init(pgd_base);
 546 }
 547
 548 pteval_t __supported_pte_mask __read_mostly = ~(_PAGE_NX | _PAGE_GLOBAL | _PAGE_IOMAP);
 549 EXPORT_SYMBOL_GPL(__supported_pte_mask);
 550
 551 /* user-defined highmem size */
 552 static unsigned int highmem_pages = -1;
 553
 554 /*
 555  * highmem=size forces highmem to be exactly 'size' bytes.
 556  * This works even on boxes that have no highmem otherwise.
 557  * This also works to reduce highmem size on bigger boxes.
 558  */
 559 static int __init parse_highmem(char *arg)
 560 {
 561         if (!arg)
 562                 return -EINVAL;
 563
 564         highmem_pages = memparse(arg, &arg) >> PAGE_SHIFT;
 565         return 0;
 566 }
 567 early_param("highmem", parse_highmem);
 568
 569 #define MSG_HIGHMEM_TOO_BIG \
 570         "highmem size (%luMB) is bigger than pages available (%luMB)!\n"
 571
 572 #define MSG_LOWMEM_TOO_SMALL \
 573         "highmem size (%luMB) results in <64MB lowmem, ignoring it!\n"
 574 /*
 575  * All of RAM fits into lowmem - but if user wants highmem
 576  * artificially via the highmem=x boot parameter then create
 577  * it:
 578  */
 579 static void __init lowmem_pfn_init(void)
 580 {
 581         /* max_low_pfn is 0, we already have early_res support */
 582         max_low_pfn = max_pfn;
 583
 584         if (highmem_pages == -1)
 585                 highmem_pages = 0;
 586 #ifdef CONFIG_HIGHMEM
 587         if (highmem_pages >= max_pfn) {
 588                 printk(KERN_ERR MSG_HIGHMEM_TOO_BIG,
 589                         pages_to_mb(highmem_pages), pages_to_mb(max_pfn));
 590                 highmem_pages = 0;
 591         }
 592         if (highmem_pages) {
 593                 if (max_low_pfn - highmem_pages < 64*1024*1024/PAGE_SIZE) {
 594                         printk(KERN_ERR MSG_LOWMEM_TOO_SMALL,
 595                                 pages_to_mb(highmem_pages));
 596                         highmem_pages = 0;
 597                 }
 598                 max_low_pfn -= highmem_pages;
 599         }
 600 #else
 601         if (highmem_pages)
 602                 printk(KERN_ERR "ignoring highmem size on non-highmem kernel!\n");
 603 #endif
 604 }
 605
 606 #define MSG_HIGHMEM_TOO_SMALL \
 607         "only %luMB highmem pages available, ignoring highmem size of %luMB!\n"
 608
 609 #define MSG_HIGHMEM_TRIMMED \
 610         "Warning: only 4GB will be used. Use a HIGHMEM64G enabled kernel!\n"
 611 /*
 612  * We have more RAM than fits into lowmem - we try to put it into
 613  * highmem, also taking the highmem=x boot parameter into account:
 614  */
 615 static void __init highmem_pfn_init(void)
 616 {
 617         max_low_pfn = MAXMEM_PFN;
 618
 619         if (highmem_pages == -1)
 620                 highmem_pages = max_pfn - MAXMEM_PFN;
 621
 622         if (highmem_pages + MAXMEM_PFN < max_pfn)
 623                 max_pfn = MAXMEM_PFN + highmem_pages;
 624
 625         if (highmem_pages + MAXMEM_PFN > max_pfn) {
 626                 printk(KERN_WARNING MSG_HIGHMEM_TOO_SMALL,
 627                         pages_to_mb(max_pfn - MAXMEM_PFN),
 628                         pages_to_mb(highmem_pages));
 629                 highmem_pages = 0;
 630         }
 631 #ifndef CONFIG_HIGHMEM
 632         /* Maximum memory usable is what is directly addressable */
 633         printk(KERN_WARNING "Warning only %ldMB will be used.\n", MAXMEM>>20);
 634         if (max_pfn > MAX_NONPAE_PFN)
 635                 printk(KERN_WARNING "Use a HIGHMEM64G enabled kernel.\n");
 636         else
 637                 printk(KERN_WARNING "Use a HIGHMEM enabled kernel.\n");
 638         max_pfn = MAXMEM_PFN;
 639 #else /* !CONFIG_HIGHMEM */
 640 #ifndef CONFIG_HIGHMEM64G
 641         if (max_pfn > MAX_NONPAE_PFN) {
 642                 max_pfn = MAX_NONPAE_PFN;
 643                 printk(KERN_WARNING MSG_HIGHMEM_TRIMMED);
 644         }
 645 #endif /* !CONFIG_HIGHMEM64G */
 646 #endif /* !CONFIG_HIGHMEM */
 647 }
 648
 649 /*
 650  * Determine low and high memory ranges:
 651  */
 652 void __init find_low_pfn_range(void)
 653 {
 654         /* it could update max_pfn */
 655
 656         if (max_pfn <= MAXMEM_PFN)
 657                 lowmem_pfn_init();
 658         else
 659                 highmem_pfn_init();
 660 }
 661
 662 #ifndef CONFIG_NEED_MULTIPLE_NODES
 663 void __init initmem_init(void)
 664 {
 665 #ifdef CONFIG_HIGHMEM
 666         highstart_pfn = highend_pfn = max_pfn;
 667         if (max_pfn > max_low_pfn)
 668                 highstart_pfn = max_low_pfn;
 669         printk(KERN_NOTICE "%ldMB HIGHMEM available.\n",
 670                 pages_to_mb(highend_pfn - highstart_pfn));
 671         num_physpages = highend_pfn;
 672         high_memory = (void *) __va(highstart_pfn * PAGE_SIZE - 1) + 1;
 673 #else
 674         num_physpages = max_low_pfn;
 675         high_memory = (void *) __va(max_low_pfn * PAGE_SIZE - 1) + 1;
 676 #endif
 677
 678         memblock_set_node(0, (phys_addr_t)ULLONG_MAX, 0);
 679         sparse_memory_present_with_active_regions(0);
 680
 681 #ifdef CONFIG_FLATMEM
 682         max_mapnr = num_physpages;
 683 #endif
 684         __vmalloc_start_set = true;
 685
 686         printk(KERN_NOTICE "%ldMB LOWMEM available.\n",
 687                         pages_to_mb(max_low_pfn));
 688
 689         setup_bootmem_allocator();
 690 }
 691 #endif /* !CONFIG_NEED_MULTIPLE_NODES */
 692
 693 void __init setup_bootmem_allocator(void)
 694 {
 695         printk(KERN_INFO "  mapped low ram: 0 - %08lx\n",
 696                  max_pfn_mapped<<PAGE_SHIFT);
 697         printk(KERN_INFO "  low ram: 0 - %08lx\n", max_low_pfn<<PAGE_SHIFT);
 698 }
 699
 700 /*
 701  * paging_init() sets up the page tables - note that the first 8MB are
 702  * already mapped by head.S.
 703  *
 704  * This routines also unmaps the page at virtual kernel address 0, so
 705  * that we can trap those pesky NULL-reference errors in the kernel.
 706  */
 707 void __init paging_init(void)
 708 {
 709         pagetable_init();
 710
 711         __flush_tlb_all();
 712
 713         kmap_init();
 714
 715         /*
 716          * NOTE: at this point the bootmem allocator is fully available.
 717          */
 718         olpc_dt_build_devicetree();
 719         sparse_memory_present_with_active_regions(MAX_NUMNODES);
 720         sparse_init();
 721         zone_sizes_init();
 722 }
 723
 724 /*
 725  * Test if the WP bit works in supervisor mode. It isn't supported on 386's
 726  * and also on some strange 486's. All 586+'s are OK. This used to involve
 727  * black magic jumps to work around some nasty CPU bugs, but fortunately the
 728  * switch to using exceptions got rid of all that.
 729  */
 730 static void __init test_wp_bit(void)
 731 {
 732         printk(KERN_INFO
 733   "Checking if this processor honours the WP bit even in supervisor mode...");
 734
 735         /* Any page-aligned address will do, the test is non-destructive */
 736         __set_fixmap(FIX_WP_TEST, __pa(&swapper_pg_dir), PAGE_KERNEL_RO);
 737         boot_cpu_data.wp_works_ok = do_test_wp_bit();
 738         clear_fixmap(FIX_WP_TEST);
 739
 740         if (!boot_cpu_data.wp_works_ok) {
 741                 printk(KERN_CONT "No.\n");
 742                 panic("Linux doesn't support CPUs with broken WP.");
 743         } else {
 744                 printk(KERN_CONT "Ok.\n");
 745         }
 746 }
 747
 748 void __init mem_init(void)
 749 {
 750         int codesize, reservedpages, datasize, initsize;
 751         int tmp;
 752
 753         pci_iommu_alloc();
 754
 755 #ifdef CONFIG_FLATMEM
 756         BUG_ON(!mem_map);
 757 #endif
 758         /*
 759          * With CONFIG_DEBUG_PAGEALLOC initialization of highmem pages has to
 760          * be done before free_all_bootmem(). Memblock use free low memory for
 761          * temporary data (see find_range_array()) and for this purpose can use
 762          * pages that was already passed to the buddy allocator, hence marked as
 763          * not accessible in the page tables when compiled with
 764          * CONFIG_DEBUG_PAGEALLOC. Otherwise order of initialization is not
 765          * important here.
 766          */
 767         set_highmem_pages_init();
 768
 769         /* this will put all low memory onto the freelists */
 770         totalram_pages += free_all_bootmem();
 771
 772         reservedpages = 0;
 773         for (tmp = 0; tmp < max_low_pfn; tmp++)
 774                 /*
 775                  * Only count reserved RAM pages:
 776                  */
 777                 if (page_is_ram(tmp) && PageReserved(pfn_to_page(tmp)))
 778                         reservedpages++;
 779
 780         after_bootmem = 1;
 781
 782         codesize =  (unsigned long) &_etext - (unsigned long) &_text;
 783         datasize =  (unsigned long) &_edata - (unsigned long) &_etext;
 784         initsize =  (unsigned long) &__init_end - (unsigned long) &__init_begin;
 785
 786         printk(KERN_INFO "Memory: %luk/%luk available (%dk kernel code, "
 787                         "%dk reserved, %dk data, %dk init, %ldk highmem)\n",
 788                 nr_free_pages() << (PAGE_SHIFT-10),
 789                 num_physpages << (PAGE_SHIFT-10),
 790                 codesize >> 10,
 791                 reservedpages << (PAGE_SHIFT-10),
 792                 datasize >> 10,
 793                 initsize >> 10,
 794                 totalhigh_pages << (PAGE_SHIFT-10));
 795
 796         printk(KERN_INFO "virtual kernel memory layout:\n"
 797                 "    fixmap  : 0x%08lx - 0x%08lx   (%4ld kB)\n"
 798 #ifdef CONFIG_HIGHMEM
 799                 "    pkmap   : 0x%08lx - 0x%08lx   (%4ld kB)\n"
 800 #endif
 801                 "    vmalloc : 0x%08lx - 0x%08lx   (%4ld MB)\n"
 802                 "    lowmem  : 0x%08lx - 0x%08lx   (%4ld MB)\n"
 803                 "      .init : 0x%08lx - 0x%08lx   (%4ld kB)\n"
 804                 "      .data : 0x%08lx - 0x%08lx   (%4ld kB)\n"
 805                 "      .text : 0x%08lx - 0x%08lx   (%4ld kB)\n",
 806                 FIXADDR_START, FIXADDR_TOP,
 807                 (FIXADDR_TOP - FIXADDR_START) >> 10,
 808
 809 #ifdef CONFIG_HIGHMEM
 810                 PKMAP_BASE, PKMAP_BASE+LAST_PKMAP*PAGE_SIZE,
 811                 (LAST_PKMAP*PAGE_SIZE) >> 10,
 812 #endif
 813
 814                 VMALLOC_START, VMALLOC_END,
 815                 (VMALLOC_END - VMALLOC_START) >> 20,
 816
 817                 (unsigned long)__va(0), (unsigned long)high_memory,
 818                 ((unsigned long)high_memory - (unsigned long)__va(0)) >> 20,
 819
 820                 (unsigned long)&__init_begin, (unsigned long)&__init_end,
 821                 ((unsigned long)&__init_end -
 822                  (unsigned long)&__init_begin) >> 10,
 823
 824                 (unsigned long)&_etext, (unsigned long)&_edata,
 825                 ((unsigned long)&_edata - (unsigned long)&_etext) >> 10,
 826
 827                 (unsigned long)&_text, (unsigned long)&_etext,
 828                 ((unsigned long)&_etext - (unsigned long)&_text) >> 10);
 829
 830         /*
 831          * Check boundaries twice: Some fundamental inconsistencies can
 832          * be detected at build time already.
 833          */
 834 #define __FIXADDR_TOP (-PAGE_SIZE)
 835 #ifdef CONFIG_HIGHMEM
 836         BUILD_BUG_ON(PKMAP_BASE + LAST_PKMAP*PAGE_SIZE  > FIXADDR_START);
 837         BUILD_BUG_ON(VMALLOC_END                        > PKMAP_BASE);
 838 #endif
 839 #define high_memory (-128UL << 20)
 840         BUILD_BUG_ON(VMALLOC_START                      >= VMALLOC_END);
 841 #undef high_memory
 842 #undef __FIXADDR_TOP
 843
 844 #ifdef CONFIG_HIGHMEM
 845         BUG_ON(PKMAP_BASE + LAST_PKMAP*PAGE_SIZE        > FIXADDR_START);
 846         BUG_ON(VMALLOC_END                              > PKMAP_BASE);
 847 #endif
 848         BUG_ON(VMALLOC_START                            >= VMALLOC_END);
 849         BUG_ON((unsigned long)high_memory               > VMALLOC_START);
 850
 851         if (boot_cpu_data.wp_works_ok < 0)
 852                 test_wp_bit();
 853 }
 854
 855 #ifdef CONFIG_MEMORY_HOTPLUG
 856 int arch_add_memory(int nid, u64 start, u64 size)
 857 {
 858         struct pglist_data *pgdata = NODE_DATA(nid);
 859         struct zone *zone = pgdata->node_zones + ZONE_HIGHMEM;
 860         unsigned long start_pfn = start >> PAGE_SHIFT;
 861         unsigned long nr_pages = size >> PAGE_SHIFT;
 862
 863         return __add_pages(nid, zone, start_pfn, nr_pages);
 864 }
 865 #endif
 866
 867 /*
 868  * This function cannot be __init, since exceptions don't work in that
 869  * section.  Put this after the callers, so that it cannot be inlined.
 870  */
 871 static noinline int do_test_wp_bit(void)
 872 {
 873         char tmp_reg;
 874         int flag;
 875
 876         __asm__ __volatile__(
 877                 "       movb %0, %1     \n"
 878                 "1:     movb %1, %0     \n"
 879                 "       xorl %2, %2     \n"
 880                 "2:                     \n"
 881                 _ASM_EXTABLE(1b,2b)
 882                 :"=m" (*(char *)fix_to_virt(FIX_WP_TEST)),
 883                  "=q" (tmp_reg),
 884                  "=r" (flag)
 885                 :"2" (1)
 886                 :"memory");
 887
 888         return flag;
 889 }
 890
 891 #ifdef CONFIG_DEBUG_RODATA
 892 const int rodata_test_data = 0xC3;
 893 EXPORT_SYMBOL_GPL(rodata_test_data);
 894
 895 int kernel_set_to_readonly __read_mostly;
 896
 897 void set_kernel_text_rw(void)
 898 {
 899         unsigned long start = PFN_ALIGN(_text);
 900         unsigned long size = PFN_ALIGN(_etext) - start;
 901
 902         if (!kernel_set_to_readonly)
 903                 return;
 904
 905         pr_debug("Set kernel text: %lx - %lx for read write\n",
 906                  start, start+size);
 907
 908         set_pages_rw(virt_to_page(start), size >> PAGE_SHIFT);
 909 }
 910
 911 void set_kernel_text_ro(void)
 912 {
 913         unsigned long start = PFN_ALIGN(_text);
 914         unsigned long size = PFN_ALIGN(_etext) - start;
 915
 916         if (!kernel_set_to_readonly)
 917                 return;
 918
 919         pr_debug("Set kernel text: %lx - %lx for read only\n",
 920                  start, start+size);
 921
 922         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
 923 }
 924
 925 static void mark_nxdata_nx(void)
 926 {
 927         /*
 928          * When this called, init has already been executed and released,
 929          * so everything past _etext should be NX.
 930          */
 931         unsigned long start = PFN_ALIGN(_etext);
 932         /*
 933          * This comes from is_kernel_text upper limit. Also HPAGE where used:
 934          */
 935         unsigned long size = (((unsigned long)__init_end + HPAGE_SIZE) & HPAGE_MASK) - start;
 936
 937         if (__supported_pte_mask & _PAGE_NX)
 938                 printk(KERN_INFO "NX-protecting the kernel data: %luk\n", size >> 10);
 939         set_pages_nx(virt_to_page(start), size >> PAGE_SHIFT);
 940 }
 941
 942 void mark_rodata_ro(void)
 943 {
 944         unsigned long start = PFN_ALIGN(_text);
 945         unsigned long size = PFN_ALIGN(_etext) - start;
 946
 947         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
 948         printk(KERN_INFO "Write protecting the kernel text: %luk\n",
 949                 size >> 10);
 950
 951         kernel_set_to_readonly = 1;
 952
 953 #ifdef CONFIG_CPA_DEBUG
 954         printk(KERN_INFO "Testing CPA: Reverting %lx-%lx\n",
 955                 start, start+size);
 956         set_pages_rw(virt_to_page(start), size>>PAGE_SHIFT);
 957
 958         printk(KERN_INFO "Testing CPA: write protecting again\n");
 959         set_pages_ro(virt_to_page(start), size>>PAGE_SHIFT);
 960 #endif
 961
 962         start += size;
 963         size = (unsigned long)__end_rodata - start;
 964         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
 965         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
 966                 size >> 10);
 967         rodata_test();
 968
 969 #ifdef CONFIG_CPA_DEBUG
 970         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, start + size);
 971         set_pages_rw(virt_to_page(start), size >> PAGE_SHIFT);
 972
 973         printk(KERN_INFO "Testing CPA: write protecting again\n");
 974         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
 975 #endif
 976         mark_nxdata_nx();
 977 }
 978 #endif
 979