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/signal.h>
   9 #include <linux/sched.h>
  10 #include <linux/kernel.h>
  11 #include <linux/errno.h>
  12 #include <linux/string.h>
  13 #include <linux/types.h>
  14 #include <linux/ptrace.h>
  15 #include <linux/mman.h>
  16 #include <linux/mm.h>
  17 #include <linux/hugetlb.h>
  18 #include <linux/swap.h>
  19 #include <linux/smp.h>
  20 #include <linux/init.h>
  21 #include <linux/highmem.h>
  22 #include <linux/pagemap.h>
  23 #include <linux/pci.h>
  24 #include <linux/pfn.h>
  25 #include <linux/poison.h>
  26 #include <linux/memblock.h>
  27 #include <linux/proc_fs.h>
  28 #include <linux/memory_hotplug.h>
  29 #include <linux/initrd.h>
  30 #include <linux/cpumask.h>
  31 #include <linux/gfp.h>
  32
  33 #include <asm/asm.h>
  34 #include <asm/bios_ebda.h>
  35 #include <asm/processor.h>
  36 #include <linux/uaccess.h>
  37 #include <asm/pgtable.h>
  38 #include <asm/dma.h>
  39 #include <asm/fixmap.h>
  40 #include <asm/e820/api.h>
  41 #include <asm/apic.h>
  42 #include <asm/bugs.h>
  43 #include <asm/tlb.h>
  44 #include <asm/tlbflush.h>
  45 #include <asm/olpc_ofw.h>
  46 #include <asm/pgalloc.h>
  47 #include <asm/sections.h>
  48 #include <asm/paravirt.h>
  49 #include <asm/setup.h>
  50 #include <asm/set_memory.h>
  51 #include <asm/page_types.h>
  52 #include <asm/cpu_entry_area.h>
  53 #include <asm/init.h>
  54
  55 #include "mm_internal.h"
  56
  57 unsigned long highstart_pfn, highend_pfn;
  58
  59 bool __read_mostly __vmalloc_start_set = false;
  60
  61 /*
  62  * Creates a middle page table and puts a pointer to it in the
  63  * given global directory entry. This only returns the gd entry
  64  * in non-PAE compilation mode, since the middle layer is folded.
  65  */
  66 static pmd_t * __init one_md_table_init(pgd_t *pgd)
  67 {
  68         p4d_t *p4d;
  69         pud_t *pud;
  70         pmd_t *pmd_table;
  71
  72 #ifdef CONFIG_X86_PAE
  73         if (!(pgd_val(*pgd) & _PAGE_PRESENT)) {
  74                 pmd_table = (pmd_t *)alloc_low_page();
  75                 paravirt_alloc_pmd(&init_mm, __pa(pmd_table) >> PAGE_SHIFT);
  76                 set_pgd(pgd, __pgd(__pa(pmd_table) | _PAGE_PRESENT));
  77                 p4d = p4d_offset(pgd, 0);
  78                 pud = pud_offset(p4d, 0);
  79                 BUG_ON(pmd_table != pmd_offset(pud, 0));
  80
  81                 return pmd_table;
  82         }
  83 #endif
  84         p4d = p4d_offset(pgd, 0);
  85         pud = pud_offset(p4d, 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         pmd_idx = pmd_index(vaddr);
 141
 142         for ( ; (pgd_idx < PTRS_PER_PGD) && (vaddr != end); pgd_idx++) {
 143                 for (; (pmd_idx < PTRS_PER_PMD) && (vaddr != end);
 144                                                         pmd_idx++) {
 145                         if ((vaddr >> PMD_SHIFT) >= pmd_idx_kmap_begin &&
 146                             (vaddr >> PMD_SHIFT) <= pmd_idx_kmap_end)
 147                                 count++;
 148                         vaddr += PMD_SIZE;
 149                 }
 150                 pmd_idx = 0;
 151         }
 152 #endif
 153         return count;
 154 }
 155
 156 static pte_t *__init page_table_kmap_check(pte_t *pte, pmd_t *pmd,
 157                                            unsigned long vaddr, pte_t *lastpte,
 158                                            void **adr)
 159 {
 160 #ifdef CONFIG_HIGHMEM
 161         /*
 162          * Something (early fixmap) may already have put a pte
 163          * page here, which causes the page table allocation
 164          * to become nonlinear. Attempt to fix it, and if it
 165          * is still nonlinear then we have to bug.
 166          */
 167         int pmd_idx_kmap_begin = fix_to_virt(FIX_KMAP_END) >> PMD_SHIFT;
 168         int pmd_idx_kmap_end = fix_to_virt(FIX_KMAP_BEGIN) >> PMD_SHIFT;
 169
 170         if (pmd_idx_kmap_begin != pmd_idx_kmap_end
 171             && (vaddr >> PMD_SHIFT) >= pmd_idx_kmap_begin
 172             && (vaddr >> PMD_SHIFT) <= pmd_idx_kmap_end) {
 173                 pte_t *newpte;
 174                 int i;
 175
 176                 BUG_ON(after_bootmem);
 177                 newpte = *adr;
 178                 for (i = 0; i < PTRS_PER_PTE; i++)
 179                         set_pte(newpte + i, pte[i]);
 180                 *adr = (void *)(((unsigned long)(*adr)) + PAGE_SIZE);
 181
 182                 paravirt_alloc_pte(&init_mm, __pa(newpte) >> PAGE_SHIFT);
 183                 set_pmd(pmd, __pmd(__pa(newpte)|_PAGE_TABLE));
 184                 BUG_ON(newpte != pte_offset_kernel(pmd, 0));
 185                 __flush_tlb_all();
 186
 187                 paravirt_release_pte(__pa(pte) >> PAGE_SHIFT);
 188                 pte = newpte;
 189         }
 190         BUG_ON(vaddr < fix_to_virt(FIX_KMAP_BEGIN - 1)
 191                && vaddr > fix_to_virt(FIX_KMAP_END)
 192                && lastpte && lastpte + PTRS_PER_PTE != pte);
 193 #endif
 194         return pte;
 195 }
 196
 197 /*
 198  * This function initializes a certain range of kernel virtual memory
 199  * with new bootmem page tables, everywhere page tables are missing in
 200  * the given range.
 201  *
 202  * NOTE: The pagetables are allocated contiguous on the physical space
 203  * so we can cache the place of the first one and move around without
 204  * checking the pgd every time.
 205  */
 206 static void __init
 207 page_table_range_init(unsigned long start, unsigned long end, pgd_t *pgd_base)
 208 {
 209         int pgd_idx, pmd_idx;
 210         unsigned long vaddr;
 211         pgd_t *pgd;
 212         pmd_t *pmd;
 213         pte_t *pte = NULL;
 214         unsigned long count = page_table_range_init_count(start, end);
 215         void *adr = NULL;
 216
 217         if (count)
 218                 adr = alloc_low_pages(count);
 219
 220         vaddr = start;
 221         pgd_idx = pgd_index(vaddr);
 222         pmd_idx = pmd_index(vaddr);
 223         pgd = pgd_base + pgd_idx;
 224
 225         for ( ; (pgd_idx < PTRS_PER_PGD) && (vaddr != end); pgd++, pgd_idx++) {
 226                 pmd = one_md_table_init(pgd);
 227                 pmd = pmd + pmd_index(vaddr);
 228                 for (; (pmd_idx < PTRS_PER_PMD) && (vaddr != end);
 229                                                         pmd++, pmd_idx++) {
 230                         pte = page_table_kmap_check(one_page_table_init(pmd),
 231                                                     pmd, vaddr, pte, &adr);
 232
 233                         vaddr += PMD_SIZE;
 234                 }
 235                 pmd_idx = 0;
 236         }
 237 }
 238
 239 static inline int is_kernel_text(unsigned long addr)
 240 {
 241         if (addr >= (unsigned long)_text && addr <= (unsigned long)__init_end)
 242                 return 1;
 243         return 0;
 244 }
 245
 246 /*
 247  * This maps the physical memory to kernel virtual address space, a total
 248  * of max_low_pfn pages, by creating page tables starting from address
 249  * PAGE_OFFSET:
 250  */
 251 unsigned long __init
 252 kernel_physical_mapping_init(unsigned long start,
 253                              unsigned long end,
 254                              unsigned long page_size_mask)
 255 {
 256         int use_pse = page_size_mask == (1<<PG_LEVEL_2M);
 257         unsigned long last_map_addr = end;
 258         unsigned long start_pfn, end_pfn;
 259         pgd_t *pgd_base = swapper_pg_dir;
 260         int pgd_idx, pmd_idx, pte_ofs;
 261         unsigned long pfn;
 262         pgd_t *pgd;
 263         pmd_t *pmd;
 264         pte_t *pte;
 265         unsigned pages_2m, pages_4k;
 266         int mapping_iter;
 267
 268         start_pfn = start >> PAGE_SHIFT;
 269         end_pfn = end >> PAGE_SHIFT;
 270
 271         /*
 272          * First iteration will setup identity mapping using large/small pages
 273          * based on use_pse, with other attributes same as set by
 274          * the early code in head_32.S
 275          *
 276          * Second iteration will setup the appropriate attributes (NX, GLOBAL..)
 277          * as desired for the kernel identity mapping.
 278          *
 279          * This two pass mechanism conforms to the TLB app note which says:
 280          *
 281          *     "Software should not write to a paging-structure entry in a way
 282          *      that would change, for any linear address, both the page size
 283          *      and either the page frame or attributes."
 284          */
 285         mapping_iter = 1;
 286
 287         if (!boot_cpu_has(X86_FEATURE_PSE))
 288                 use_pse = 0;
 289
 290 repeat:
 291         pages_2m = pages_4k = 0;
 292         pfn = start_pfn;
 293         pgd_idx = pgd_index((pfn<<PAGE_SHIFT) + PAGE_OFFSET);
 294         pgd = pgd_base + pgd_idx;
 295         for (; pgd_idx < PTRS_PER_PGD; pgd++, pgd_idx++) {
 296                 pmd = one_md_table_init(pgd);
 297
 298                 if (pfn >= end_pfn)
 299                         continue;
 300 #ifdef CONFIG_X86_PAE
 301                 pmd_idx = pmd_index((pfn<<PAGE_SHIFT) + PAGE_OFFSET);
 302                 pmd += pmd_idx;
 303 #else
 304                 pmd_idx = 0;
 305 #endif
 306                 for (; pmd_idx < PTRS_PER_PMD && pfn < end_pfn;
 307                      pmd++, pmd_idx++) {
 308                         unsigned int addr = pfn * PAGE_SIZE + PAGE_OFFSET;
 309
 310                         /*
 311                          * Map with big pages if possible, otherwise
 312                          * create normal page tables:
 313                          */
 314                         if (use_pse) {
 315                                 unsigned int addr2;
 316                                 pgprot_t prot = PAGE_KERNEL_LARGE;
 317                                 /*
 318                                  * first pass will use the same initial
 319                                  * identity mapping attribute + _PAGE_PSE.
 320                                  */
 321                                 pgprot_t init_prot =
 322                                         __pgprot(PTE_IDENT_ATTR |
 323                                                  _PAGE_PSE);
 324
 325                                 pfn &= PMD_MASK >> PAGE_SHIFT;
 326                                 addr2 = (pfn + PTRS_PER_PTE-1) * PAGE_SIZE +
 327                                         PAGE_OFFSET + PAGE_SIZE-1;
 328
 329                                 if (is_kernel_text(addr) ||
 330                                     is_kernel_text(addr2))
 331                                         prot = PAGE_KERNEL_LARGE_EXEC;
 332
 333                                 pages_2m++;
 334                                 if (mapping_iter == 1)
 335                                         set_pmd(pmd, pfn_pmd(pfn, init_prot));
 336                                 else
 337                                         set_pmd(pmd, pfn_pmd(pfn, prot));
 338
 339                                 pfn += PTRS_PER_PTE;
 340                                 continue;
 341                         }
 342                         pte = one_page_table_init(pmd);
 343
 344                         pte_ofs = pte_index((pfn<<PAGE_SHIFT) + PAGE_OFFSET);
 345                         pte += pte_ofs;
 346                         for (; pte_ofs < PTRS_PER_PTE && pfn < end_pfn;
 347                              pte++, pfn++, pte_ofs++, addr += PAGE_SIZE) {
 348                                 pgprot_t prot = PAGE_KERNEL;
 349                                 /*
 350                                  * first pass will use the same initial
 351                                  * identity mapping attribute.
 352                                  */
 353                                 pgprot_t init_prot = __pgprot(PTE_IDENT_ATTR);
 354
 355                                 if (is_kernel_text(addr))
 356                                         prot = PAGE_KERNEL_EXEC;
 357
 358                                 pages_4k++;
 359                                 if (mapping_iter == 1) {
 360                                         set_pte(pte, pfn_pte(pfn, init_prot));
 361                                         last_map_addr = (pfn << PAGE_SHIFT) + PAGE_SIZE;
 362                                 } else
 363                                         set_pte(pte, pfn_pte(pfn, prot));
 364                         }
 365                 }
 366         }
 367         if (mapping_iter == 1) {
 368                 /*
 369                  * update direct mapping page count only in the first
 370                  * iteration.
 371                  */
 372                 update_page_count(PG_LEVEL_2M, pages_2m);
 373                 update_page_count(PG_LEVEL_4K, pages_4k);
 374
 375                 /*
 376                  * local global flush tlb, which will flush the previous
 377                  * mappings present in both small and large page TLB's.
 378                  */
 379                 __flush_tlb_all();
 380
 381                 /*
 382                  * Second iteration will set the actual desired PTE attributes.
 383                  */
 384                 mapping_iter = 2;
 385                 goto repeat;
 386         }
 387         return last_map_addr;
 388 }
 389
 390 pte_t *kmap_pte;
 391
 392 static inline pte_t *kmap_get_fixmap_pte(unsigned long vaddr)
 393 {
 394         pgd_t *pgd = pgd_offset_k(vaddr);
 395         p4d_t *p4d = p4d_offset(pgd, vaddr);
 396         pud_t *pud = pud_offset(p4d, vaddr);
 397         pmd_t *pmd = pmd_offset(pud, vaddr);
 398         return pte_offset_kernel(pmd, vaddr);
 399 }
 400
 401 static void __init kmap_init(void)
 402 {
 403         unsigned long kmap_vstart;
 404
 405         /*
 406          * Cache the first kmap pte:
 407          */
 408         kmap_vstart = __fix_to_virt(FIX_KMAP_BEGIN);
 409         kmap_pte = kmap_get_fixmap_pte(kmap_vstart);
 410 }
 411
 412 #ifdef CONFIG_HIGHMEM
 413 static void __init permanent_kmaps_init(pgd_t *pgd_base)
 414 {
 415         unsigned long vaddr;
 416         pgd_t *pgd;
 417         p4d_t *p4d;
 418         pud_t *pud;
 419         pmd_t *pmd;
 420         pte_t *pte;
 421
 422         vaddr = PKMAP_BASE;
 423         page_table_range_init(vaddr, vaddr + PAGE_SIZE*LAST_PKMAP, pgd_base);
 424
 425         pgd = swapper_pg_dir + pgd_index(vaddr);
 426         p4d = p4d_offset(pgd, vaddr);
 427         pud = pud_offset(p4d, vaddr);
 428         pmd = pmd_offset(pud, vaddr);
 429         pte = pte_offset_kernel(pmd, vaddr);
 430         pkmap_page_table = pte;
 431 }
 432
 433 void __init add_highpages_with_active_regions(int nid,
 434                          unsigned long start_pfn, unsigned long end_pfn)
 435 {
 436         phys_addr_t start, end;
 437         u64 i;
 438
 439         for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &start, &end, NULL) {
 440                 unsigned long pfn = clamp_t(unsigned long, PFN_UP(start),
 441                                             start_pfn, end_pfn);
 442                 unsigned long e_pfn = clamp_t(unsigned long, PFN_DOWN(end),
 443                                               start_pfn, end_pfn);
 444                 for ( ; pfn < e_pfn; pfn++)
 445                         if (pfn_valid(pfn))
 446                                 free_highmem_page(pfn_to_page(pfn));
 447         }
 448 }
 449 #else
 450 static inline void permanent_kmaps_init(pgd_t *pgd_base)
 451 {
 452 }
 453 #endif /* CONFIG_HIGHMEM */
 454
 455 void __init sync_initial_page_table(void)
 456 {
 457         clone_pgd_range(initial_page_table + KERNEL_PGD_BOUNDARY,
 458                         swapper_pg_dir     + KERNEL_PGD_BOUNDARY,
 459                         KERNEL_PGD_PTRS);
 460
 461         /*
 462          * sync back low identity map too.  It is used for example
 463          * in the 32-bit EFI stub.
 464          */
 465         clone_pgd_range(initial_page_table,
 466                         swapper_pg_dir     + KERNEL_PGD_BOUNDARY,
 467                         min(KERNEL_PGD_PTRS, KERNEL_PGD_BOUNDARY));
 468 }
 469
 470 void __init native_pagetable_init(void)
 471 {
 472         unsigned long pfn, va;
 473         pgd_t *pgd, *base = swapper_pg_dir;
 474         p4d_t *p4d;
 475         pud_t *pud;
 476         pmd_t *pmd;
 477         pte_t *pte;
 478
 479         /*
 480          * Remove any mappings which extend past the end of physical
 481          * memory from the boot time page table.
 482          * In virtual address space, we should have at least two pages
 483          * from VMALLOC_END to pkmap or fixmap according to VMALLOC_END
 484          * definition. And max_low_pfn is set to VMALLOC_END physical
 485          * address. If initial memory mapping is doing right job, we
 486          * should have pte used near max_low_pfn or one pmd is not present.
 487          */
 488         for (pfn = max_low_pfn; pfn < 1<<(32-PAGE_SHIFT); pfn++) {
 489                 va = PAGE_OFFSET + (pfn<<PAGE_SHIFT);
 490                 pgd = base + pgd_index(va);
 491                 if (!pgd_present(*pgd))
 492                         break;
 493
 494                 p4d = p4d_offset(pgd, va);
 495                 pud = pud_offset(p4d, va);
 496                 pmd = pmd_offset(pud, va);
 497                 if (!pmd_present(*pmd))
 498                         break;
 499
 500                 /* should not be large page here */
 501                 if (pmd_large(*pmd)) {
 502                         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",
 503                                 pfn, pmd, __pa(pmd));
 504                         BUG_ON(1);
 505                 }
 506
 507                 pte = pte_offset_kernel(pmd, va);
 508                 if (!pte_present(*pte))
 509                         break;
 510
 511                 printk(KERN_DEBUG "clearing pte for ram above max_low_pfn: pfn: %lx pmd: %p pmd phys: %lx pte: %p pte phys: %lx\n",
 512                                 pfn, pmd, __pa(pmd), pte, __pa(pte));
 513                 pte_clear(NULL, va, pte);
 514         }
 515         paravirt_alloc_pmd(&init_mm, __pa(base) >> PAGE_SHIFT);
 516         paging_init();
 517 }
 518
 519 /*
 520  * Build a proper pagetable for the kernel mappings.  Up until this
 521  * point, we've been running on some set of pagetables constructed by
 522  * the boot process.
 523  *
 524  * If we're booting on native hardware, this will be a pagetable
 525  * constructed in arch/x86/kernel/head_32.S.  The root of the
 526  * pagetable will be swapper_pg_dir.
 527  *
 528  * If we're booting paravirtualized under a hypervisor, then there are
 529  * more options: we may already be running PAE, and the pagetable may
 530  * or may not be based in swapper_pg_dir.  In any case,
 531  * paravirt_pagetable_init() will set up swapper_pg_dir
 532  * appropriately for the rest of the initialization to work.
 533  *
 534  * In general, pagetable_init() assumes that the pagetable may already
 535  * be partially populated, and so it avoids stomping on any existing
 536  * mappings.
 537  */
 538 void __init early_ioremap_page_table_range_init(void)
 539 {
 540         pgd_t *pgd_base = swapper_pg_dir;
 541         unsigned long vaddr, end;
 542
 543         /*
 544          * Fixed mappings, only the page table structure has to be
 545          * created - mappings will be set by set_fixmap():
 546          */
 547         vaddr = __fix_to_virt(__end_of_fixed_addresses - 1) & PMD_MASK;
 548         end = (FIXADDR_TOP + PMD_SIZE - 1) & PMD_MASK;
 549         page_table_range_init(vaddr, end, pgd_base);
 550         early_ioremap_reset();
 551 }
 552
 553 static void __init pagetable_init(void)
 554 {
 555         pgd_t *pgd_base = swapper_pg_dir;
 556
 557         permanent_kmaps_init(pgd_base);
 558 }
 559
 560 #define DEFAULT_PTE_MASK ~(_PAGE_NX | _PAGE_GLOBAL)
 561 /* Bits supported by the hardware: */
 562 pteval_t __supported_pte_mask __read_mostly = DEFAULT_PTE_MASK;
 563 /* Bits allowed in normal kernel mappings: */
 564 pteval_t __default_kernel_pte_mask __read_mostly = DEFAULT_PTE_MASK;
 565 EXPORT_SYMBOL_GPL(__supported_pte_mask);
 566 /* Used in PAGE_KERNEL_* macros which are reasonably used out-of-tree: */
 567 EXPORT_SYMBOL(__default_kernel_pte_mask);
 568
 569 /* user-defined highmem size */
 570 static unsigned int highmem_pages = -1;
 571
 572 /*
 573  * highmem=size forces highmem to be exactly 'size' bytes.
 574  * This works even on boxes that have no highmem otherwise.
 575  * This also works to reduce highmem size on bigger boxes.
 576  */
 577 static int __init parse_highmem(char *arg)
 578 {
 579         if (!arg)
 580                 return -EINVAL;
 581
 582         highmem_pages = memparse(arg, &arg) >> PAGE_SHIFT;
 583         return 0;
 584 }
 585 early_param("highmem", parse_highmem);
 586
 587 #define MSG_HIGHMEM_TOO_BIG \
 588         "highmem size (%luMB) is bigger than pages available (%luMB)!\n"
 589
 590 #define MSG_LOWMEM_TOO_SMALL \
 591         "highmem size (%luMB) results in <64MB lowmem, ignoring it!\n"
 592 /*
 593  * All of RAM fits into lowmem - but if user wants highmem
 594  * artificially via the highmem=x boot parameter then create
 595  * it:
 596  */
 597 static void __init lowmem_pfn_init(void)
 598 {
 599         /* max_low_pfn is 0, we already have early_res support */
 600         max_low_pfn = max_pfn;
 601
 602         if (highmem_pages == -1)
 603                 highmem_pages = 0;
 604 #ifdef CONFIG_HIGHMEM
 605         if (highmem_pages >= max_pfn) {
 606                 printk(KERN_ERR MSG_HIGHMEM_TOO_BIG,
 607                         pages_to_mb(highmem_pages), pages_to_mb(max_pfn));
 608                 highmem_pages = 0;
 609         }
 610         if (highmem_pages) {
 611                 if (max_low_pfn - highmem_pages < 64*1024*1024/PAGE_SIZE) {
 612                         printk(KERN_ERR MSG_LOWMEM_TOO_SMALL,
 613                                 pages_to_mb(highmem_pages));
 614                         highmem_pages = 0;
 615                 }
 616                 max_low_pfn -= highmem_pages;
 617         }
 618 #else
 619         if (highmem_pages)
 620                 printk(KERN_ERR "ignoring highmem size on non-highmem kernel!\n");
 621 #endif
 622 }
 623
 624 #define MSG_HIGHMEM_TOO_SMALL \
 625         "only %luMB highmem pages available, ignoring highmem size of %luMB!\n"
 626
 627 #define MSG_HIGHMEM_TRIMMED \
 628         "Warning: only 4GB will be used. Use a HIGHMEM64G enabled kernel!\n"
 629 /*
 630  * We have more RAM than fits into lowmem - we try to put it into
 631  * highmem, also taking the highmem=x boot parameter into account:
 632  */
 633 static void __init highmem_pfn_init(void)
 634 {
 635         max_low_pfn = MAXMEM_PFN;
 636
 637         if (highmem_pages == -1)
 638                 highmem_pages = max_pfn - MAXMEM_PFN;
 639
 640         if (highmem_pages + MAXMEM_PFN < max_pfn)
 641                 max_pfn = MAXMEM_PFN + highmem_pages;
 642
 643         if (highmem_pages + MAXMEM_PFN > max_pfn) {
 644                 printk(KERN_WARNING MSG_HIGHMEM_TOO_SMALL,
 645                         pages_to_mb(max_pfn - MAXMEM_PFN),
 646                         pages_to_mb(highmem_pages));
 647                 highmem_pages = 0;
 648         }
 649 #ifndef CONFIG_HIGHMEM
 650         /* Maximum memory usable is what is directly addressable */
 651         printk(KERN_WARNING "Warning only %ldMB will be used.\n", MAXMEM>>20);
 652         if (max_pfn > MAX_NONPAE_PFN)
 653                 printk(KERN_WARNING "Use a HIGHMEM64G enabled kernel.\n");
 654         else
 655                 printk(KERN_WARNING "Use a HIGHMEM enabled kernel.\n");
 656         max_pfn = MAXMEM_PFN;
 657 #else /* !CONFIG_HIGHMEM */
 658 #ifndef CONFIG_HIGHMEM64G
 659         if (max_pfn > MAX_NONPAE_PFN) {
 660                 max_pfn = MAX_NONPAE_PFN;
 661                 printk(KERN_WARNING MSG_HIGHMEM_TRIMMED);
 662         }
 663 #endif /* !CONFIG_HIGHMEM64G */
 664 #endif /* !CONFIG_HIGHMEM */
 665 }
 666
 667 /*
 668  * Determine low and high memory ranges:
 669  */
 670 void __init find_low_pfn_range(void)
 671 {
 672         /* it could update max_pfn */
 673
 674         if (max_pfn <= MAXMEM_PFN)
 675                 lowmem_pfn_init();
 676         else
 677                 highmem_pfn_init();
 678 }
 679
 680 #ifndef CONFIG_NEED_MULTIPLE_NODES
 681 void __init initmem_init(void)
 682 {
 683 #ifdef CONFIG_HIGHMEM
 684         highstart_pfn = highend_pfn = max_pfn;
 685         if (max_pfn > max_low_pfn)
 686                 highstart_pfn = max_low_pfn;
 687         printk(KERN_NOTICE "%ldMB HIGHMEM available.\n",
 688                 pages_to_mb(highend_pfn - highstart_pfn));
 689         high_memory = (void *) __va(highstart_pfn * PAGE_SIZE - 1) + 1;
 690 #else
 691         high_memory = (void *) __va(max_low_pfn * PAGE_SIZE - 1) + 1;
 692 #endif
 693
 694         memblock_set_node(0, PHYS_ADDR_MAX, &memblock.memory, 0);
 695         sparse_memory_present_with_active_regions(0);
 696
 697 #ifdef CONFIG_FLATMEM
 698         max_mapnr = IS_ENABLED(CONFIG_HIGHMEM) ? highend_pfn : max_low_pfn;
 699 #endif
 700         __vmalloc_start_set = true;
 701
 702         printk(KERN_NOTICE "%ldMB LOWMEM available.\n",
 703                         pages_to_mb(max_low_pfn));
 704
 705         setup_bootmem_allocator();
 706 }
 707 #endif /* !CONFIG_NEED_MULTIPLE_NODES */
 708
 709 void __init setup_bootmem_allocator(void)
 710 {
 711         printk(KERN_INFO "  mapped low ram: 0 - %08lx\n",
 712                  max_pfn_mapped<<PAGE_SHIFT);
 713         printk(KERN_INFO "  low ram: 0 - %08lx\n", max_low_pfn<<PAGE_SHIFT);
 714 }
 715
 716 /*
 717  * paging_init() sets up the page tables - note that the first 8MB are
 718  * already mapped by head.S.
 719  *
 720  * This routines also unmaps the page at virtual kernel address 0, so
 721  * that we can trap those pesky NULL-reference errors in the kernel.
 722  */
 723 void __init paging_init(void)
 724 {
 725         pagetable_init();
 726
 727         __flush_tlb_all();
 728
 729         kmap_init();
 730
 731         /*
 732          * NOTE: at this point the bootmem allocator is fully available.
 733          */
 734         olpc_dt_build_devicetree();
 735         sparse_memory_present_with_active_regions(MAX_NUMNODES);
 736         sparse_init();
 737         zone_sizes_init();
 738 }
 739
 740 /*
 741  * Test if the WP bit works in supervisor mode. It isn't supported on 386's
 742  * and also on some strange 486's. All 586+'s are OK. This used to involve
 743  * black magic jumps to work around some nasty CPU bugs, but fortunately the
 744  * switch to using exceptions got rid of all that.
 745  */
 746 static void __init test_wp_bit(void)
 747 {
 748         char z = 0;
 749
 750         printk(KERN_INFO "Checking if this processor honours the WP bit even in supervisor mode...");
 751
 752         __set_fixmap(FIX_WP_TEST, __pa_symbol(empty_zero_page), PAGE_KERNEL_RO);
 753
 754         if (probe_kernel_write((char *)fix_to_virt(FIX_WP_TEST), &z, 1)) {
 755                 clear_fixmap(FIX_WP_TEST);
 756                 printk(KERN_CONT "Ok.\n");
 757                 return;
 758         }
 759
 760         printk(KERN_CONT "No.\n");
 761         panic("Linux doesn't support CPUs with broken WP.");
 762 }
 763
 764 void __init mem_init(void)
 765 {
 766         pci_iommu_alloc();
 767
 768 #ifdef CONFIG_FLATMEM
 769         BUG_ON(!mem_map);
 770 #endif
 771         /*
 772          * With CONFIG_DEBUG_PAGEALLOC initialization of highmem pages has to
 773          * be done before memblock_free_all(). Memblock use free low memory for
 774          * temporary data (see find_range_array()) and for this purpose can use
 775          * pages that was already passed to the buddy allocator, hence marked as
 776          * not accessible in the page tables when compiled with
 777          * CONFIG_DEBUG_PAGEALLOC. Otherwise order of initialization is not
 778          * important here.
 779          */
 780         set_highmem_pages_init();
 781
 782         /* this will put all low memory onto the freelists */
 783         memblock_free_all();
 784
 785         after_bootmem = 1;
 786         x86_init.hyper.init_after_bootmem();
 787
 788         mem_init_print_info(NULL);
 789         printk(KERN_INFO "virtual kernel memory layout:\n"
 790                 "    fixmap  : 0x%08lx - 0x%08lx   (%4ld kB)\n"
 791                 "  cpu_entry : 0x%08lx - 0x%08lx   (%4ld kB)\n"
 792 #ifdef CONFIG_HIGHMEM
 793                 "    pkmap   : 0x%08lx - 0x%08lx   (%4ld kB)\n"
 794 #endif
 795                 "    vmalloc : 0x%08lx - 0x%08lx   (%4ld MB)\n"
 796                 "    lowmem  : 0x%08lx - 0x%08lx   (%4ld MB)\n"
 797                 "      .init : 0x%08lx - 0x%08lx   (%4ld kB)\n"
 798                 "      .data : 0x%08lx - 0x%08lx   (%4ld kB)\n"
 799                 "      .text : 0x%08lx - 0x%08lx   (%4ld kB)\n",
 800                 FIXADDR_START, FIXADDR_TOP,
 801                 (FIXADDR_TOP - FIXADDR_START) >> 10,
 802
 803                 CPU_ENTRY_AREA_BASE,
 804                 CPU_ENTRY_AREA_BASE + CPU_ENTRY_AREA_MAP_SIZE,
 805                 CPU_ENTRY_AREA_MAP_SIZE >> 10,
 806
 807 #ifdef CONFIG_HIGHMEM
 808                 PKMAP_BASE, PKMAP_BASE+LAST_PKMAP*PAGE_SIZE,
 809                 (LAST_PKMAP*PAGE_SIZE) >> 10,
 810 #endif
 811
 812                 VMALLOC_START, VMALLOC_END,
 813                 (VMALLOC_END - VMALLOC_START) >> 20,
 814
 815                 (unsigned long)__va(0), (unsigned long)high_memory,
 816                 ((unsigned long)high_memory - (unsigned long)__va(0)) >> 20,
 817
 818                 (unsigned long)&__init_begin, (unsigned long)&__init_end,
 819                 ((unsigned long)&__init_end -
 820                  (unsigned long)&__init_begin) >> 10,
 821
 822                 (unsigned long)&_etext, (unsigned long)&_edata,
 823                 ((unsigned long)&_edata - (unsigned long)&_etext) >> 10,
 824
 825                 (unsigned long)&_text, (unsigned long)&_etext,
 826                 ((unsigned long)&_etext - (unsigned long)&_text) >> 10);
 827
 828         /*
 829          * Check boundaries twice: Some fundamental inconsistencies can
 830          * be detected at build time already.
 831          */
 832 #define __FIXADDR_TOP (-PAGE_SIZE)
 833 #ifdef CONFIG_HIGHMEM
 834         BUILD_BUG_ON(PKMAP_BASE + LAST_PKMAP*PAGE_SIZE  > FIXADDR_START);
 835         BUILD_BUG_ON(VMALLOC_END                        > PKMAP_BASE);
 836 #endif
 837 #define high_memory (-128UL << 20)
 838         BUILD_BUG_ON(VMALLOC_START                      >= VMALLOC_END);
 839 #undef high_memory
 840 #undef __FIXADDR_TOP
 841
 842 #ifdef CONFIG_HIGHMEM
 843         BUG_ON(PKMAP_BASE + LAST_PKMAP*PAGE_SIZE        > FIXADDR_START);
 844         BUG_ON(VMALLOC_END                              > PKMAP_BASE);
 845 #endif
 846         BUG_ON(VMALLOC_START                            >= VMALLOC_END);
 847         BUG_ON((unsigned long)high_memory               > VMALLOC_START);
 848
 849         test_wp_bit();
 850 }
 851
 852 #ifdef CONFIG_MEMORY_HOTPLUG
 853 int arch_add_memory(int nid, u64 start, u64 size, struct vmem_altmap *altmap,
 854                 bool want_memblock)
 855 {
 856         unsigned long start_pfn = start >> PAGE_SHIFT;
 857         unsigned long nr_pages = size >> PAGE_SHIFT;
 858
 859         return __add_pages(nid, start_pfn, nr_pages, altmap, want_memblock);
 860 }
 861
 862 #ifdef CONFIG_MEMORY_HOTREMOVE
 863 int arch_remove_memory(int nid, u64 start, u64 size, struct vmem_altmap *altmap)
 864 {
 865         unsigned long start_pfn = start >> PAGE_SHIFT;
 866         unsigned long nr_pages = size >> PAGE_SHIFT;
 867         struct zone *zone;
 868
 869         zone = page_zone(pfn_to_page(start_pfn));
 870         return __remove_pages(zone, start_pfn, nr_pages, altmap);
 871 }
 872 #endif
 873 #endif
 874
 875 int kernel_set_to_readonly __read_mostly;
 876
 877 void set_kernel_text_rw(void)
 878 {
 879         unsigned long start = PFN_ALIGN(_text);
 880         unsigned long size = PFN_ALIGN(_etext) - start;
 881
 882         if (!kernel_set_to_readonly)
 883                 return;
 884
 885         pr_debug("Set kernel text: %lx - %lx for read write\n",
 886                  start, start+size);
 887
 888         set_pages_rw(virt_to_page(start), size >> PAGE_SHIFT);
 889 }
 890
 891 void set_kernel_text_ro(void)
 892 {
 893         unsigned long start = PFN_ALIGN(_text);
 894         unsigned long size = PFN_ALIGN(_etext) - start;
 895
 896         if (!kernel_set_to_readonly)
 897                 return;
 898
 899         pr_debug("Set kernel text: %lx - %lx for read only\n",
 900                  start, start+size);
 901
 902         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
 903 }
 904
 905 static void mark_nxdata_nx(void)
 906 {
 907         /*
 908          * When this called, init has already been executed and released,
 909          * so everything past _etext should be NX.
 910          */
 911         unsigned long start = PFN_ALIGN(_etext);
 912         /*
 913          * This comes from is_kernel_text upper limit. Also HPAGE where used:
 914          */
 915         unsigned long size = (((unsigned long)__init_end + HPAGE_SIZE) & HPAGE_MASK) - start;
 916
 917         if (__supported_pte_mask & _PAGE_NX)
 918                 printk(KERN_INFO "NX-protecting the kernel data: %luk\n", size >> 10);
 919         set_pages_nx(virt_to_page(start), size >> PAGE_SHIFT);
 920 }
 921
 922 void mark_rodata_ro(void)
 923 {
 924         unsigned long start = PFN_ALIGN(_text);
 925         unsigned long size = (unsigned long)__end_rodata - start;
 926
 927         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
 928         pr_info("Write protecting kernel text and read-only data: %luk\n",
 929                 size >> 10);
 930
 931         kernel_set_to_readonly = 1;
 932
 933 #ifdef CONFIG_CPA_DEBUG
 934         pr_info("Testing CPA: Reverting %lx-%lx\n", start, start + size);
 935         set_pages_rw(virt_to_page(start), size >> PAGE_SHIFT);
 936
 937         pr_info("Testing CPA: write protecting again\n");
 938         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
 939 #endif
 940         mark_nxdata_nx();
 941         if (__supported_pte_mask & _PAGE_NX)
 942                 debug_checkwx();
 943 }