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/bootmem.h>
  27 #include <linux/memblock.h>
  28 #include <linux/proc_fs.h>
  29 #include <linux/memory_hotplug.h>
  30 #include <linux/initrd.h>
  31 #include <linux/cpumask.h>
  32 #include <linux/gfp.h>
  33
  34 #include <asm/asm.h>
  35 #include <asm/bios_ebda.h>
  36 #include <asm/processor.h>
  37 #include <linux/uaccess.h>
  38 #include <asm/pgtable.h>
  39 #include <asm/dma.h>
  40 #include <asm/fixmap.h>
  41 #include <asm/e820.h>
  42 #include <asm/apic.h>
  43 #include <asm/bugs.h>
  44 #include <asm/tlb.h>
  45 #include <asm/tlbflush.h>
  46 #include <asm/olpc_ofw.h>
  47 #include <asm/pgalloc.h>
  48 #include <asm/sections.h>
  49 #include <asm/paravirt.h>
  50 #include <asm/setup.h>
  51 #include <asm/cacheflush.h>
  52 #include <asm/page_types.h>
  53 #include <asm/init.h>
  54
  55 #include "mm_internal.h"
  56
  57 unsigned long highstart_pfn, highend_pfn;
  58
  59 static noinline int do_test_wp_bit(void);
  60
  61 bool __read_mostly __vmalloc_start_set = false;
  62
  63 /*
  64  * Creates a middle page table and puts a pointer to it in the
  65  * given global directory entry. This only returns the gd entry
  66  * in non-PAE compilation mode, since the middle layer is folded.
  67  */
  68 static pmd_t * __init one_md_table_init(pgd_t *pgd)
  69 {
  70         p4d_t *p4d;
  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                 p4d = p4d_offset(pgd, 0);
  80                 pud = pud_offset(p4d, 0);
  81                 BUG_ON(pmd_table != pmd_offset(pud, 0));
  82
  83                 return pmd_table;
  84         }
  85 #endif
  86         p4d = p4d_offset(pgd, 0);
  87         pud = pud_offset(p4d, 0);
  88         pmd_table = pmd_offset(pud, 0);
  89
  90         return pmd_table;
  91 }
  92
  93 /*
  94  * Create a page table and place a pointer to it in a middle page
  95  * directory entry:
  96  */
  97 static pte_t * __init one_page_table_init(pmd_t *pmd)
  98 {
  99         if (!(pmd_val(*pmd) & _PAGE_PRESENT)) {
 100                 pte_t *page_table = (pte_t *)alloc_low_page();
 101
 102                 paravirt_alloc_pte(&init_mm, __pa(page_table) >> PAGE_SHIFT);
 103                 set_pmd(pmd, __pmd(__pa(page_table) | _PAGE_TABLE));
 104                 BUG_ON(page_table != pte_offset_kernel(pmd, 0));
 105         }
 106
 107         return pte_offset_kernel(pmd, 0);
 108 }
 109
 110 pmd_t * __init populate_extra_pmd(unsigned long vaddr)
 111 {
 112         int pgd_idx = pgd_index(vaddr);
 113         int pmd_idx = pmd_index(vaddr);
 114
 115         return one_md_table_init(swapper_pg_dir + pgd_idx) + pmd_idx;
 116 }
 117
 118 pte_t * __init populate_extra_pte(unsigned long vaddr)
 119 {
 120         int pte_idx = pte_index(vaddr);
 121         pmd_t *pmd;
 122
 123         pmd = populate_extra_pmd(vaddr);
 124         return one_page_table_init(pmd) + pte_idx;
 125 }
 126
 127 static unsigned long __init
 128 page_table_range_init_count(unsigned long start, unsigned long end)
 129 {
 130         unsigned long count = 0;
 131 #ifdef CONFIG_HIGHMEM
 132         int pmd_idx_kmap_begin = fix_to_virt(FIX_KMAP_END) >> PMD_SHIFT;
 133         int pmd_idx_kmap_end = fix_to_virt(FIX_KMAP_BEGIN) >> PMD_SHIFT;
 134         int pgd_idx, pmd_idx;
 135         unsigned long vaddr;
 136
 137         if (pmd_idx_kmap_begin == pmd_idx_kmap_end)
 138                 return 0;
 139
 140         vaddr = start;
 141         pgd_idx = pgd_index(vaddr);
 142         pmd_idx = pmd_index(vaddr);
 143
 144         for ( ; (pgd_idx < PTRS_PER_PGD) && (vaddr != end); pgd_idx++) {
 145                 for (; (pmd_idx < PTRS_PER_PMD) && (vaddr != end);
 146                                                         pmd_idx++) {
 147                         if ((vaddr >> PMD_SHIFT) >= pmd_idx_kmap_begin &&
 148                             (vaddr >> PMD_SHIFT) <= pmd_idx_kmap_end)
 149                                 count++;
 150                         vaddr += PMD_SIZE;
 151                 }
 152                 pmd_idx = 0;
 153         }
 154 #endif
 155         return count;
 156 }
 157
 158 static pte_t *__init page_table_kmap_check(pte_t *pte, pmd_t *pmd,
 159                                            unsigned long vaddr, pte_t *lastpte,
 160                                            void **adr)
 161 {
 162 #ifdef CONFIG_HIGHMEM
 163         /*
 164          * Something (early fixmap) may already have put a pte
 165          * page here, which causes the page table allocation
 166          * to become nonlinear. Attempt to fix it, and if it
 167          * is still nonlinear then we have to bug.
 168          */
 169         int pmd_idx_kmap_begin = fix_to_virt(FIX_KMAP_END) >> PMD_SHIFT;
 170         int pmd_idx_kmap_end = fix_to_virt(FIX_KMAP_BEGIN) >> PMD_SHIFT;
 171
 172         if (pmd_idx_kmap_begin != pmd_idx_kmap_end
 173             && (vaddr >> PMD_SHIFT) >= pmd_idx_kmap_begin
 174             && (vaddr >> PMD_SHIFT) <= pmd_idx_kmap_end) {
 175                 pte_t *newpte;
 176                 int i;
 177
 178                 BUG_ON(after_bootmem);
 179                 newpte = *adr;
 180                 for (i = 0; i < PTRS_PER_PTE; i++)
 181                         set_pte(newpte + i, pte[i]);
 182                 *adr = (void *)(((unsigned long)(*adr)) + PAGE_SIZE);
 183
 184                 paravirt_alloc_pte(&init_mm, __pa(newpte) >> PAGE_SHIFT);
 185                 set_pmd(pmd, __pmd(__pa(newpte)|_PAGE_TABLE));
 186                 BUG_ON(newpte != pte_offset_kernel(pmd, 0));
 187                 __flush_tlb_all();
 188
 189                 paravirt_release_pte(__pa(pte) >> PAGE_SHIFT);
 190                 pte = newpte;
 191         }
 192         BUG_ON(vaddr < fix_to_virt(FIX_KMAP_BEGIN - 1)
 193                && vaddr > fix_to_virt(FIX_KMAP_END)
 194                && lastpte && lastpte + PTRS_PER_PTE != pte);
 195 #endif
 196         return pte;
 197 }
 198
 199 /*
 200  * This function initializes a certain range of kernel virtual memory
 201  * with new bootmem page tables, everywhere page tables are missing in
 202  * the given range.
 203  *
 204  * NOTE: The pagetables are allocated contiguous on the physical space
 205  * so we can cache the place of the first one and move around without
 206  * checking the pgd every time.
 207  */
 208 static void __init
 209 page_table_range_init(unsigned long start, unsigned long end, pgd_t *pgd_base)
 210 {
 211         int pgd_idx, pmd_idx;
 212         unsigned long vaddr;
 213         pgd_t *pgd;
 214         pmd_t *pmd;
 215         pte_t *pte = NULL;
 216         unsigned long count = page_table_range_init_count(start, end);
 217         void *adr = NULL;
 218
 219         if (count)
 220                 adr = alloc_low_pages(count);
 221
 222         vaddr = start;
 223         pgd_idx = pgd_index(vaddr);
 224         pmd_idx = pmd_index(vaddr);
 225         pgd = pgd_base + pgd_idx;
 226
 227         for ( ; (pgd_idx < PTRS_PER_PGD) && (vaddr != end); pgd++, pgd_idx++) {
 228                 pmd = one_md_table_init(pgd);
 229                 pmd = pmd + pmd_index(vaddr);
 230                 for (; (pmd_idx < PTRS_PER_PMD) && (vaddr != end);
 231                                                         pmd++, pmd_idx++) {
 232                         pte = page_table_kmap_check(one_page_table_init(pmd),
 233                                                     pmd, vaddr, pte, &adr);
 234
 235                         vaddr += PMD_SIZE;
 236                 }
 237                 pmd_idx = 0;
 238         }
 239 }
 240
 241 static inline int is_kernel_text(unsigned long addr)
 242 {
 243         if (addr >= (unsigned long)_text && addr <= (unsigned long)__init_end)
 244                 return 1;
 245         return 0;
 246 }
 247
 248 /*
 249  * This maps the physical memory to kernel virtual address space, a total
 250  * of max_low_pfn pages, by creating page tables starting from address
 251  * PAGE_OFFSET:
 252  */
 253 unsigned long __init
 254 kernel_physical_mapping_init(unsigned long start,
 255                              unsigned long end,
 256                              unsigned long page_size_mask)
 257 {
 258         int use_pse = page_size_mask == (1<<PG_LEVEL_2M);
 259         unsigned long last_map_addr = end;
 260         unsigned long start_pfn, end_pfn;
 261         pgd_t *pgd_base = swapper_pg_dir;
 262         int pgd_idx, pmd_idx, pte_ofs;
 263         unsigned long pfn;
 264         pgd_t *pgd;
 265         pmd_t *pmd;
 266         pte_t *pte;
 267         unsigned pages_2m, pages_4k;
 268         int mapping_iter;
 269
 270         start_pfn = start >> PAGE_SHIFT;
 271         end_pfn = end >> PAGE_SHIFT;
 272
 273         /*
 274          * First iteration will setup identity mapping using large/small pages
 275          * based on use_pse, with other attributes same as set by
 276          * the early code in head_32.S
 277          *
 278          * Second iteration will setup the appropriate attributes (NX, GLOBAL..)
 279          * as desired for the kernel identity mapping.
 280          *
 281          * This two pass mechanism conforms to the TLB app note which says:
 282          *
 283          *     "Software should not write to a paging-structure entry in a way
 284          *      that would change, for any linear address, both the page size
 285          *      and either the page frame or attributes."
 286          */
 287         mapping_iter = 1;
 288
 289         if (!boot_cpu_has(X86_FEATURE_PSE))
 290                 use_pse = 0;
 291
 292 repeat:
 293         pages_2m = pages_4k = 0;
 294         pfn = start_pfn;
 295         pgd_idx = pgd_index((pfn<<PAGE_SHIFT) + PAGE_OFFSET);
 296         pgd = pgd_base + pgd_idx;
 297         for (; pgd_idx < PTRS_PER_PGD; pgd++, pgd_idx++) {
 298                 pmd = one_md_table_init(pgd);
 299
 300                 if (pfn >= end_pfn)
 301                         continue;
 302 #ifdef CONFIG_X86_PAE
 303                 pmd_idx = pmd_index((pfn<<PAGE_SHIFT) + PAGE_OFFSET);
 304                 pmd += pmd_idx;
 305 #else
 306                 pmd_idx = 0;
 307 #endif
 308                 for (; pmd_idx < PTRS_PER_PMD && pfn < end_pfn;
 309                      pmd++, pmd_idx++) {
 310                         unsigned int addr = pfn * PAGE_SIZE + PAGE_OFFSET;
 311
 312                         /*
 313                          * Map with big pages if possible, otherwise
 314                          * create normal page tables:
 315                          */
 316                         if (use_pse) {
 317                                 unsigned int addr2;
 318                                 pgprot_t prot = PAGE_KERNEL_LARGE;
 319                                 /*
 320                                  * first pass will use the same initial
 321                                  * identity mapping attribute + _PAGE_PSE.
 322                                  */
 323                                 pgprot_t init_prot =
 324                                         __pgprot(PTE_IDENT_ATTR |
 325                                                  _PAGE_PSE);
 326
 327                                 pfn &= PMD_MASK >> PAGE_SHIFT;
 328                                 addr2 = (pfn + PTRS_PER_PTE-1) * PAGE_SIZE +
 329                                         PAGE_OFFSET + PAGE_SIZE-1;
 330
 331                                 if (is_kernel_text(addr) ||
 332                                     is_kernel_text(addr2))
 333                                         prot = PAGE_KERNEL_LARGE_EXEC;
 334
 335                                 pages_2m++;
 336                                 if (mapping_iter == 1)
 337                                         set_pmd(pmd, pfn_pmd(pfn, init_prot));
 338                                 else
 339                                         set_pmd(pmd, pfn_pmd(pfn, prot));
 340
 341                                 pfn += PTRS_PER_PTE;
 342                                 continue;
 343                         }
 344                         pte = one_page_table_init(pmd);
 345
 346                         pte_ofs = pte_index((pfn<<PAGE_SHIFT) + PAGE_OFFSET);
 347                         pte += pte_ofs;
 348                         for (; pte_ofs < PTRS_PER_PTE && pfn < end_pfn;
 349                              pte++, pfn++, pte_ofs++, addr += PAGE_SIZE) {
 350                                 pgprot_t prot = PAGE_KERNEL;
 351                                 /*
 352                                  * first pass will use the same initial
 353                                  * identity mapping attribute.
 354                                  */
 355                                 pgprot_t init_prot = __pgprot(PTE_IDENT_ATTR);
 356
 357                                 if (is_kernel_text(addr))
 358                                         prot = PAGE_KERNEL_EXEC;
 359
 360                                 pages_4k++;
 361                                 if (mapping_iter == 1) {
 362                                         set_pte(pte, pfn_pte(pfn, init_prot));
 363                                         last_map_addr = (pfn << PAGE_SHIFT) + PAGE_SIZE;
 364                                 } else
 365                                         set_pte(pte, pfn_pte(pfn, prot));
 366                         }
 367                 }
 368         }
 369         if (mapping_iter == 1) {
 370                 /*
 371                  * update direct mapping page count only in the first
 372                  * iteration.
 373                  */
 374                 update_page_count(PG_LEVEL_2M, pages_2m);
 375                 update_page_count(PG_LEVEL_4K, pages_4k);
 376
 377                 /*
 378                  * local global flush tlb, which will flush the previous
 379                  * mappings present in both small and large page TLB's.
 380                  */
 381                 __flush_tlb_all();
 382
 383                 /*
 384                  * Second iteration will set the actual desired PTE attributes.
 385                  */
 386                 mapping_iter = 2;
 387                 goto repeat;
 388         }
 389         return last_map_addr;
 390 }
 391
 392 pte_t *kmap_pte;
 393
 394 static inline pte_t *kmap_get_fixmap_pte(unsigned long vaddr)
 395 {
 396         pgd_t *pgd = pgd_offset_k(vaddr);
 397         p4d_t *p4d = p4d_offset(pgd, vaddr);
 398         pud_t *pud = pud_offset(p4d, vaddr);
 399         pmd_t *pmd = pmd_offset(pud, vaddr);
 400         return pte_offset_kernel(pmd, vaddr);
 401 }
 402
 403 static void __init kmap_init(void)
 404 {
 405         unsigned long kmap_vstart;
 406
 407         /*
 408          * Cache the first kmap pte:
 409          */
 410         kmap_vstart = __fix_to_virt(FIX_KMAP_BEGIN);
 411         kmap_pte = kmap_get_fixmap_pte(kmap_vstart);
 412 }
 413
 414 #ifdef CONFIG_HIGHMEM
 415 static void __init permanent_kmaps_init(pgd_t *pgd_base)
 416 {
 417         unsigned long vaddr;
 418         pgd_t *pgd;
 419         p4d_t *p4d;
 420         pud_t *pud;
 421         pmd_t *pmd;
 422         pte_t *pte;
 423
 424         vaddr = PKMAP_BASE;
 425         page_table_range_init(vaddr, vaddr + PAGE_SIZE*LAST_PKMAP, pgd_base);
 426
 427         pgd = swapper_pg_dir + pgd_index(vaddr);
 428         p4d = p4d_offset(pgd, vaddr);
 429         pud = pud_offset(p4d, vaddr);
 430         pmd = pmd_offset(pud, vaddr);
 431         pte = pte_offset_kernel(pmd, vaddr);
 432         pkmap_page_table = pte;
 433 }
 434
 435 void __init add_highpages_with_active_regions(int nid,
 436                          unsigned long start_pfn, unsigned long end_pfn)
 437 {
 438         phys_addr_t start, end;
 439         u64 i;
 440
 441         for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &start, &end, NULL) {
 442                 unsigned long pfn = clamp_t(unsigned long, PFN_UP(start),
 443                                             start_pfn, end_pfn);
 444                 unsigned long e_pfn = clamp_t(unsigned long, PFN_DOWN(end),
 445                                               start_pfn, end_pfn);
 446                 for ( ; pfn < e_pfn; pfn++)
 447                         if (pfn_valid(pfn))
 448                                 free_highmem_page(pfn_to_page(pfn));
 449         }
 450 }
 451 #else
 452 static inline void permanent_kmaps_init(pgd_t *pgd_base)
 453 {
 454 }
 455 #endif /* CONFIG_HIGHMEM */
 456
 457 void __init native_pagetable_init(void)
 458 {
 459         unsigned long pfn, va;
 460         pgd_t *pgd, *base = swapper_pg_dir;
 461         p4d_t *p4d;
 462         pud_t *pud;
 463         pmd_t *pmd;
 464         pte_t *pte;
 465
 466         /*
 467          * Remove any mappings which extend past the end of physical
 468          * memory from the boot time page table.
 469          * In virtual address space, we should have at least two pages
 470          * from VMALLOC_END to pkmap or fixmap according to VMALLOC_END
 471          * definition. And max_low_pfn is set to VMALLOC_END physical
 472          * address. If initial memory mapping is doing right job, we
 473          * should have pte used near max_low_pfn or one pmd is not present.
 474          */
 475         for (pfn = max_low_pfn; pfn < 1<<(32-PAGE_SHIFT); pfn++) {
 476                 va = PAGE_OFFSET + (pfn<<PAGE_SHIFT);
 477                 pgd = base + pgd_index(va);
 478                 if (!pgd_present(*pgd))
 479                         break;
 480
 481                 p4d = p4d_offset(pgd, va);
 482                 pud = pud_offset(p4d, va);
 483                 pmd = pmd_offset(pud, va);
 484                 if (!pmd_present(*pmd))
 485                         break;
 486
 487                 /* should not be large page here */
 488                 if (pmd_large(*pmd)) {
 489                         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",
 490                                 pfn, pmd, __pa(pmd));
 491                         BUG_ON(1);
 492                 }
 493
 494                 pte = pte_offset_kernel(pmd, va);
 495                 if (!pte_present(*pte))
 496                         break;
 497
 498                 printk(KERN_DEBUG "clearing pte for ram above max_low_pfn: pfn: %lx pmd: %p pmd phys: %lx pte: %p pte phys: %lx\n",
 499                                 pfn, pmd, __pa(pmd), pte, __pa(pte));
 500                 pte_clear(NULL, va, pte);
 501         }
 502         paravirt_alloc_pmd(&init_mm, __pa(base) >> PAGE_SHIFT);
 503         paging_init();
 504 }
 505
 506 /*
 507  * Build a proper pagetable for the kernel mappings.  Up until this
 508  * point, we've been running on some set of pagetables constructed by
 509  * the boot process.
 510  *
 511  * If we're booting on native hardware, this will be a pagetable
 512  * constructed in arch/x86/kernel/head_32.S.  The root of the
 513  * pagetable will be swapper_pg_dir.
 514  *
 515  * If we're booting paravirtualized under a hypervisor, then there are
 516  * more options: we may already be running PAE, and the pagetable may
 517  * or may not be based in swapper_pg_dir.  In any case,
 518  * paravirt_pagetable_init() will set up swapper_pg_dir
 519  * appropriately for the rest of the initialization to work.
 520  *
 521  * In general, pagetable_init() assumes that the pagetable may already
 522  * be partially populated, and so it avoids stomping on any existing
 523  * mappings.
 524  */
 525 void __init early_ioremap_page_table_range_init(void)
 526 {
 527         pgd_t *pgd_base = swapper_pg_dir;
 528         unsigned long vaddr, end;
 529
 530         /*
 531          * Fixed mappings, only the page table structure has to be
 532          * created - mappings will be set by set_fixmap():
 533          */
 534         vaddr = __fix_to_virt(__end_of_fixed_addresses - 1) & PMD_MASK;
 535         end = (FIXADDR_TOP + PMD_SIZE - 1) & PMD_MASK;
 536         page_table_range_init(vaddr, end, pgd_base);
 537         early_ioremap_reset();
 538 }
 539
 540 static void __init pagetable_init(void)
 541 {
 542         pgd_t *pgd_base = swapper_pg_dir;
 543
 544         permanent_kmaps_init(pgd_base);
 545 }
 546
 547 pteval_t __supported_pte_mask __read_mostly = ~(_PAGE_NX | _PAGE_GLOBAL);
 548 EXPORT_SYMBOL_GPL(__supported_pte_mask);
 549
 550 /* user-defined highmem size */
 551 static unsigned int highmem_pages = -1;
 552
 553 /*
 554  * highmem=size forces highmem to be exactly 'size' bytes.
 555  * This works even on boxes that have no highmem otherwise.
 556  * This also works to reduce highmem size on bigger boxes.
 557  */
 558 static int __init parse_highmem(char *arg)
 559 {
 560         if (!arg)
 561                 return -EINVAL;
 562
 563         highmem_pages = memparse(arg, &arg) >> PAGE_SHIFT;
 564         return 0;
 565 }
 566 early_param("highmem", parse_highmem);
 567
 568 #define MSG_HIGHMEM_TOO_BIG \
 569         "highmem size (%luMB) is bigger than pages available (%luMB)!\n"
 570
 571 #define MSG_LOWMEM_TOO_SMALL \
 572         "highmem size (%luMB) results in <64MB lowmem, ignoring it!\n"
 573 /*
 574  * All of RAM fits into lowmem - but if user wants highmem
 575  * artificially via the highmem=x boot parameter then create
 576  * it:
 577  */
 578 static void __init lowmem_pfn_init(void)
 579 {
 580         /* max_low_pfn is 0, we already have early_res support */
 581         max_low_pfn = max_pfn;
 582
 583         if (highmem_pages == -1)
 584                 highmem_pages = 0;
 585 #ifdef CONFIG_HIGHMEM
 586         if (highmem_pages >= max_pfn) {
 587                 printk(KERN_ERR MSG_HIGHMEM_TOO_BIG,
 588                         pages_to_mb(highmem_pages), pages_to_mb(max_pfn));
 589                 highmem_pages = 0;
 590         }
 591         if (highmem_pages) {
 592                 if (max_low_pfn - highmem_pages < 64*1024*1024/PAGE_SIZE) {
 593                         printk(KERN_ERR MSG_LOWMEM_TOO_SMALL,
 594                                 pages_to_mb(highmem_pages));
 595                         highmem_pages = 0;
 596                 }
 597                 max_low_pfn -= highmem_pages;
 598         }
 599 #else
 600         if (highmem_pages)
 601                 printk(KERN_ERR "ignoring highmem size on non-highmem kernel!\n");
 602 #endif
 603 }
 604
 605 #define MSG_HIGHMEM_TOO_SMALL \
 606         "only %luMB highmem pages available, ignoring highmem size of %luMB!\n"
 607
 608 #define MSG_HIGHMEM_TRIMMED \
 609         "Warning: only 4GB will be used. Use a HIGHMEM64G enabled kernel!\n"
 610 /*
 611  * We have more RAM than fits into lowmem - we try to put it into
 612  * highmem, also taking the highmem=x boot parameter into account:
 613  */
 614 static void __init highmem_pfn_init(void)
 615 {
 616         max_low_pfn = MAXMEM_PFN;
 617
 618         if (highmem_pages == -1)
 619                 highmem_pages = max_pfn - MAXMEM_PFN;
 620
 621         if (highmem_pages + MAXMEM_PFN < max_pfn)
 622                 max_pfn = MAXMEM_PFN + highmem_pages;
 623
 624         if (highmem_pages + MAXMEM_PFN > max_pfn) {
 625                 printk(KERN_WARNING MSG_HIGHMEM_TOO_SMALL,
 626                         pages_to_mb(max_pfn - MAXMEM_PFN),
 627                         pages_to_mb(highmem_pages));
 628                 highmem_pages = 0;
 629         }
 630 #ifndef CONFIG_HIGHMEM
 631         /* Maximum memory usable is what is directly addressable */
 632         printk(KERN_WARNING "Warning only %ldMB will be used.\n", MAXMEM>>20);
 633         if (max_pfn > MAX_NONPAE_PFN)
 634                 printk(KERN_WARNING "Use a HIGHMEM64G enabled kernel.\n");
 635         else
 636                 printk(KERN_WARNING "Use a HIGHMEM enabled kernel.\n");
 637         max_pfn = MAXMEM_PFN;
 638 #else /* !CONFIG_HIGHMEM */
 639 #ifndef CONFIG_HIGHMEM64G
 640         if (max_pfn > MAX_NONPAE_PFN) {
 641                 max_pfn = MAX_NONPAE_PFN;
 642                 printk(KERN_WARNING MSG_HIGHMEM_TRIMMED);
 643         }
 644 #endif /* !CONFIG_HIGHMEM64G */
 645 #endif /* !CONFIG_HIGHMEM */
 646 }
 647
 648 /*
 649  * Determine low and high memory ranges:
 650  */
 651 void __init find_low_pfn_range(void)
 652 {
 653         /* it could update max_pfn */
 654
 655         if (max_pfn <= MAXMEM_PFN)
 656                 lowmem_pfn_init();
 657         else
 658                 highmem_pfn_init();
 659 }
 660
 661 #ifndef CONFIG_NEED_MULTIPLE_NODES
 662 void __init initmem_init(void)
 663 {
 664 #ifdef CONFIG_HIGHMEM
 665         highstart_pfn = highend_pfn = max_pfn;
 666         if (max_pfn > max_low_pfn)
 667                 highstart_pfn = max_low_pfn;
 668         printk(KERN_NOTICE "%ldMB HIGHMEM available.\n",
 669                 pages_to_mb(highend_pfn - highstart_pfn));
 670         high_memory = (void *) __va(highstart_pfn * PAGE_SIZE - 1) + 1;
 671 #else
 672         high_memory = (void *) __va(max_low_pfn * PAGE_SIZE - 1) + 1;
 673 #endif
 674
 675         memblock_set_node(0, (phys_addr_t)ULLONG_MAX, &memblock.memory, 0);
 676         sparse_memory_present_with_active_regions(0);
 677
 678 #ifdef CONFIG_FLATMEM
 679         max_mapnr = IS_ENABLED(CONFIG_HIGHMEM) ? highend_pfn : max_low_pfn;
 680 #endif
 681         __vmalloc_start_set = true;
 682
 683         printk(KERN_NOTICE "%ldMB LOWMEM available.\n",
 684                         pages_to_mb(max_low_pfn));
 685
 686         setup_bootmem_allocator();
 687 }
 688 #endif /* !CONFIG_NEED_MULTIPLE_NODES */
 689
 690 void __init setup_bootmem_allocator(void)
 691 {
 692         printk(KERN_INFO "  mapped low ram: 0 - %08lx\n",
 693                  max_pfn_mapped<<PAGE_SHIFT);
 694         printk(KERN_INFO "  low ram: 0 - %08lx\n", max_low_pfn<<PAGE_SHIFT);
 695 }
 696
 697 /*
 698  * paging_init() sets up the page tables - note that the first 8MB are
 699  * already mapped by head.S.
 700  *
 701  * This routines also unmaps the page at virtual kernel address 0, so
 702  * that we can trap those pesky NULL-reference errors in the kernel.
 703  */
 704 void __init paging_init(void)
 705 {
 706         pagetable_init();
 707
 708         __flush_tlb_all();
 709
 710         kmap_init();
 711
 712         /*
 713          * NOTE: at this point the bootmem allocator is fully available.
 714          */
 715         olpc_dt_build_devicetree();
 716         sparse_memory_present_with_active_regions(MAX_NUMNODES);
 717         sparse_init();
 718         zone_sizes_init();
 719 }
 720
 721 /*
 722  * Test if the WP bit works in supervisor mode. It isn't supported on 386's
 723  * and also on some strange 486's. All 586+'s are OK. This used to involve
 724  * black magic jumps to work around some nasty CPU bugs, but fortunately the
 725  * switch to using exceptions got rid of all that.
 726  */
 727 static void __init test_wp_bit(void)
 728 {
 729         int wp_works_ok;
 730
 731         printk(KERN_INFO
 732   "Checking if this processor honours the WP bit even in supervisor mode...");
 733
 734         /* Any page-aligned address will do, the test is non-destructive */
 735         __set_fixmap(FIX_WP_TEST, __pa(&swapper_pg_dir), PAGE_KERNEL_RO);
 736         wp_works_ok = do_test_wp_bit();
 737         clear_fixmap(FIX_WP_TEST);
 738
 739         if (!wp_works_ok) {
 740                 printk(KERN_CONT "No.\n");
 741                 panic("Linux doesn't support CPUs with broken WP.");
 742         } else {
 743                 printk(KERN_CONT "Ok.\n");
 744         }
 745 }
 746
 747 void __init mem_init(void)
 748 {
 749         pci_iommu_alloc();
 750
 751 #ifdef CONFIG_FLATMEM
 752         BUG_ON(!mem_map);
 753 #endif
 754         /*
 755          * With CONFIG_DEBUG_PAGEALLOC initialization of highmem pages has to
 756          * be done before free_all_bootmem(). Memblock use free low memory for
 757          * temporary data (see find_range_array()) and for this purpose can use
 758          * pages that was already passed to the buddy allocator, hence marked as
 759          * not accessible in the page tables when compiled with
 760          * CONFIG_DEBUG_PAGEALLOC. Otherwise order of initialization is not
 761          * important here.
 762          */
 763         set_highmem_pages_init();
 764
 765         /* this will put all low memory onto the freelists */
 766         free_all_bootmem();
 767
 768         after_bootmem = 1;
 769
 770         mem_init_print_info(NULL);
 771         printk(KERN_INFO "virtual kernel memory layout:\n"
 772                 "    fixmap  : 0x%08lx - 0x%08lx   (%4ld kB)\n"
 773 #ifdef CONFIG_HIGHMEM
 774                 "    pkmap   : 0x%08lx - 0x%08lx   (%4ld kB)\n"
 775 #endif
 776                 "    vmalloc : 0x%08lx - 0x%08lx   (%4ld MB)\n"
 777                 "    lowmem  : 0x%08lx - 0x%08lx   (%4ld MB)\n"
 778                 "      .init : 0x%08lx - 0x%08lx   (%4ld kB)\n"
 779                 "      .data : 0x%08lx - 0x%08lx   (%4ld kB)\n"
 780                 "      .text : 0x%08lx - 0x%08lx   (%4ld kB)\n",
 781                 FIXADDR_START, FIXADDR_TOP,
 782                 (FIXADDR_TOP - FIXADDR_START) >> 10,
 783
 784 #ifdef CONFIG_HIGHMEM
 785                 PKMAP_BASE, PKMAP_BASE+LAST_PKMAP*PAGE_SIZE,
 786                 (LAST_PKMAP*PAGE_SIZE) >> 10,
 787 #endif
 788
 789                 VMALLOC_START, VMALLOC_END,
 790                 (VMALLOC_END - VMALLOC_START) >> 20,
 791
 792                 (unsigned long)__va(0), (unsigned long)high_memory,
 793                 ((unsigned long)high_memory - (unsigned long)__va(0)) >> 20,
 794
 795                 (unsigned long)&__init_begin, (unsigned long)&__init_end,
 796                 ((unsigned long)&__init_end -
 797                  (unsigned long)&__init_begin) >> 10,
 798
 799                 (unsigned long)&_etext, (unsigned long)&_edata,
 800                 ((unsigned long)&_edata - (unsigned long)&_etext) >> 10,
 801
 802                 (unsigned long)&_text, (unsigned long)&_etext,
 803                 ((unsigned long)&_etext - (unsigned long)&_text) >> 10);
 804
 805         /*
 806          * Check boundaries twice: Some fundamental inconsistencies can
 807          * be detected at build time already.
 808          */
 809 #define __FIXADDR_TOP (-PAGE_SIZE)
 810 #ifdef CONFIG_HIGHMEM
 811         BUILD_BUG_ON(PKMAP_BASE + LAST_PKMAP*PAGE_SIZE  > FIXADDR_START);
 812         BUILD_BUG_ON(VMALLOC_END                        > PKMAP_BASE);
 813 #endif
 814 #define high_memory (-128UL << 20)
 815         BUILD_BUG_ON(VMALLOC_START                      >= VMALLOC_END);
 816 #undef high_memory
 817 #undef __FIXADDR_TOP
 818
 819 #ifdef CONFIG_HIGHMEM
 820         BUG_ON(PKMAP_BASE + LAST_PKMAP*PAGE_SIZE        > FIXADDR_START);
 821         BUG_ON(VMALLOC_END                              > PKMAP_BASE);
 822 #endif
 823         BUG_ON(VMALLOC_START                            >= VMALLOC_END);
 824         BUG_ON((unsigned long)high_memory               > VMALLOC_START);
 825
 826         test_wp_bit();
 827 }
 828
 829 #ifdef CONFIG_MEMORY_HOTPLUG
 830 int arch_add_memory(int nid, u64 start, u64 size, bool for_device)
 831 {
 832         struct pglist_data *pgdata = NODE_DATA(nid);
 833         struct zone *zone = pgdata->node_zones +
 834                 zone_for_memory(nid, start, size, ZONE_HIGHMEM, for_device);
 835         unsigned long start_pfn = start >> PAGE_SHIFT;
 836         unsigned long nr_pages = size >> PAGE_SHIFT;
 837
 838         return __add_pages(nid, zone, start_pfn, nr_pages);
 839 }
 840
 841 #ifdef CONFIG_MEMORY_HOTREMOVE
 842 int arch_remove_memory(u64 start, u64 size)
 843 {
 844         unsigned long start_pfn = start >> PAGE_SHIFT;
 845         unsigned long nr_pages = size >> PAGE_SHIFT;
 846         struct zone *zone;
 847
 848         zone = page_zone(pfn_to_page(start_pfn));
 849         return __remove_pages(zone, start_pfn, nr_pages);
 850 }
 851 #endif
 852 #endif
 853
 854 /*
 855  * This function cannot be __init, since exceptions don't work in that
 856  * section.  Put this after the callers, so that it cannot be inlined.
 857  */
 858 static noinline int do_test_wp_bit(void)
 859 {
 860         char tmp_reg;
 861         int flag;
 862
 863         __asm__ __volatile__(
 864                 "       movb %0, %1     \n"
 865                 "1:     movb %1, %0     \n"
 866                 "       xorl %2, %2     \n"
 867                 "2:                     \n"
 868                 _ASM_EXTABLE(1b,2b)
 869                 :"=m" (*(char *)fix_to_virt(FIX_WP_TEST)),
 870                  "=q" (tmp_reg),
 871                  "=r" (flag)
 872                 :"2" (1)
 873                 :"memory");
 874
 875         return flag;
 876 }
 877
 878 int kernel_set_to_readonly __read_mostly;
 879
 880 void set_kernel_text_rw(void)
 881 {
 882         unsigned long start = PFN_ALIGN(_text);
 883         unsigned long size = PFN_ALIGN(_etext) - start;
 884
 885         if (!kernel_set_to_readonly)
 886                 return;
 887
 888         pr_debug("Set kernel text: %lx - %lx for read write\n",
 889                  start, start+size);
 890
 891         set_pages_rw(virt_to_page(start), size >> PAGE_SHIFT);
 892 }
 893
 894 void set_kernel_text_ro(void)
 895 {
 896         unsigned long start = PFN_ALIGN(_text);
 897         unsigned long size = PFN_ALIGN(_etext) - start;
 898
 899         if (!kernel_set_to_readonly)
 900                 return;
 901
 902         pr_debug("Set kernel text: %lx - %lx for read only\n",
 903                  start, start+size);
 904
 905         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
 906 }
 907
 908 static void mark_nxdata_nx(void)
 909 {
 910         /*
 911          * When this called, init has already been executed and released,
 912          * so everything past _etext should be NX.
 913          */
 914         unsigned long start = PFN_ALIGN(_etext);
 915         /*
 916          * This comes from is_kernel_text upper limit. Also HPAGE where used:
 917          */
 918         unsigned long size = (((unsigned long)__init_end + HPAGE_SIZE) & HPAGE_MASK) - start;
 919
 920         if (__supported_pte_mask & _PAGE_NX)
 921                 printk(KERN_INFO "NX-protecting the kernel data: %luk\n", size >> 10);
 922         set_pages_nx(virt_to_page(start), size >> PAGE_SHIFT);
 923 }
 924
 925 void mark_rodata_ro(void)
 926 {
 927         unsigned long start = PFN_ALIGN(_text);
 928         unsigned long size = PFN_ALIGN(_etext) - start;
 929
 930         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
 931         printk(KERN_INFO "Write protecting the kernel text: %luk\n",
 932                 size >> 10);
 933
 934         kernel_set_to_readonly = 1;
 935
 936 #ifdef CONFIG_CPA_DEBUG
 937         printk(KERN_INFO "Testing CPA: Reverting %lx-%lx\n",
 938                 start, start+size);
 939         set_pages_rw(virt_to_page(start), size>>PAGE_SHIFT);
 940
 941         printk(KERN_INFO "Testing CPA: write protecting again\n");
 942         set_pages_ro(virt_to_page(start), size>>PAGE_SHIFT);
 943 #endif
 944
 945         start += size;
 946         size = (unsigned long)__end_rodata - start;
 947         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
 948         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
 949                 size >> 10);
 950
 951 #ifdef CONFIG_CPA_DEBUG
 952         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, start + size);
 953         set_pages_rw(virt_to_page(start), size >> PAGE_SHIFT);
 954
 955         printk(KERN_INFO "Testing CPA: write protecting again\n");
 956         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
 957 #endif
 958         mark_nxdata_nx();
 959         if (__supported_pte_mask & _PAGE_NX)
 960                 debug_checkwx();
 961 }