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/api.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/set_memory.h>
  52 #include <asm/page_types.h>
  53 #include <asm/cpu_entry_area.h>
  54 #include <asm/init.h>
  55
  56 #include "mm_internal.h"
  57
  58 unsigned long highstart_pfn, highend_pfn;
  59
  60 bool __read_mostly __vmalloc_start_set = false;
  61
  62 /*
  63  * Creates a middle page table and puts a pointer to it in the
  64  * given global directory entry. This only returns the gd entry
  65  * in non-PAE compilation mode, since the middle layer is folded.
  66  */
  67 static pmd_t * __init one_md_table_init(pgd_t *pgd)
  68 {
  69         p4d_t *p4d;
  70         pud_t *pud;
  71         pmd_t *pmd_table;
  72
  73 #ifdef CONFIG_X86_PAE
  74         if (!(pgd_val(*pgd) & _PAGE_PRESENT)) {
  75                 pmd_table = (pmd_t *)alloc_low_page();
  76                 paravirt_alloc_pmd(&init_mm, __pa(pmd_table) >> PAGE_SHIFT);
  77                 set_pgd(pgd, __pgd(__pa(pmd_table) | _PAGE_PRESENT));
  78                 p4d = p4d_offset(pgd, 0);
  79                 pud = pud_offset(p4d, 0);
  80                 BUG_ON(pmd_table != pmd_offset(pud, 0));
  81
  82                 return pmd_table;
  83         }
  84 #endif
  85         p4d = p4d_offset(pgd, 0);
  86         pud = pud_offset(p4d, 0);
  87         pmd_table = pmd_offset(pud, 0);
  88
  89         return pmd_table;
  90 }
  91
  92 /*
  93  * Create a page table and place a pointer to it in a middle page
  94  * directory entry:
  95  */
  96 static pte_t * __init one_page_table_init(pmd_t *pmd)
  97 {
  98         if (!(pmd_val(*pmd) & _PAGE_PRESENT)) {
  99                 pte_t *page_table = (pte_t *)alloc_low_page();
 100
 101                 paravirt_alloc_pte(&init_mm, __pa(page_table) >> PAGE_SHIFT);
 102                 set_pmd(pmd, __pmd(__pa(page_table) | _PAGE_TABLE));
 103                 BUG_ON(page_table != pte_offset_kernel(pmd, 0));
 104         }
 105
 106         return pte_offset_kernel(pmd, 0);
 107 }
 108
 109 pmd_t * __init populate_extra_pmd(unsigned long vaddr)
 110 {
 111         int pgd_idx = pgd_index(vaddr);
 112         int pmd_idx = pmd_index(vaddr);
 113
 114         return one_md_table_init(swapper_pg_dir + pgd_idx) + pmd_idx;
 115 }
 116
 117 pte_t * __init populate_extra_pte(unsigned long vaddr)
 118 {
 119         int pte_idx = pte_index(vaddr);
 120         pmd_t *pmd;
 121
 122         pmd = populate_extra_pmd(vaddr);
 123         return one_page_table_init(pmd) + pte_idx;
 124 }
 125
 126 static unsigned long __init
 127 page_table_range_init_count(unsigned long start, unsigned long end)
 128 {
 129         unsigned long count = 0;
 130 #ifdef CONFIG_HIGHMEM
 131         int pmd_idx_kmap_begin = fix_to_virt(FIX_KMAP_END) >> PMD_SHIFT;
 132         int pmd_idx_kmap_end = fix_to_virt(FIX_KMAP_BEGIN) >> PMD_SHIFT;
 133         int pgd_idx, pmd_idx;
 134         unsigned long vaddr;
 135
 136         if (pmd_idx_kmap_begin == pmd_idx_kmap_end)
 137                 return 0;
 138
 139         vaddr = start;
 140         pgd_idx = pgd_index(vaddr);
 141         pmd_idx = pmd_index(vaddr);
 142
 143         for ( ; (pgd_idx < PTRS_PER_PGD) && (vaddr != end); pgd_idx++) {
 144                 for (; (pmd_idx < PTRS_PER_PMD) && (vaddr != end);
 145                                                         pmd_idx++) {
 146                         if ((vaddr >> PMD_SHIFT) >= pmd_idx_kmap_begin &&
 147                             (vaddr >> PMD_SHIFT) <= pmd_idx_kmap_end)
 148                                 count++;
 149                         vaddr += PMD_SIZE;
 150                 }
 151                 pmd_idx = 0;
 152         }
 153 #endif
 154         return count;
 155 }
 156
 157 static pte_t *__init page_table_kmap_check(pte_t *pte, pmd_t *pmd,
 158                                            unsigned long vaddr, pte_t *lastpte,
 159                                            void **adr)
 160 {
 161 #ifdef CONFIG_HIGHMEM
 162         /*
 163          * Something (early fixmap) may already have put a pte
 164          * page here, which causes the page table allocation
 165          * to become nonlinear. Attempt to fix it, and if it
 166          * is still nonlinear then we have to bug.
 167          */
 168         int pmd_idx_kmap_begin = fix_to_virt(FIX_KMAP_END) >> PMD_SHIFT;
 169         int pmd_idx_kmap_end = fix_to_virt(FIX_KMAP_BEGIN) >> PMD_SHIFT;
 170
 171         if (pmd_idx_kmap_begin != pmd_idx_kmap_end
 172             && (vaddr >> PMD_SHIFT) >= pmd_idx_kmap_begin
 173             && (vaddr >> PMD_SHIFT) <= pmd_idx_kmap_end) {
 174                 pte_t *newpte;
 175                 int i;
 176
 177                 BUG_ON(after_bootmem);
 178                 newpte = *adr;
 179                 for (i = 0; i < PTRS_PER_PTE; i++)
 180                         set_pte(newpte + i, pte[i]);
 181                 *adr = (void *)(((unsigned long)(*adr)) + PAGE_SIZE);
 182
 183                 paravirt_alloc_pte(&init_mm, __pa(newpte) >> PAGE_SHIFT);
 184                 set_pmd(pmd, __pmd(__pa(newpte)|_PAGE_TABLE));
 185                 BUG_ON(newpte != pte_offset_kernel(pmd, 0));
 186                 __flush_tlb_all();
 187
 188                 paravirt_release_pte(__pa(pte) >> PAGE_SHIFT);
 189                 pte = newpte;
 190         }
 191         BUG_ON(vaddr < fix_to_virt(FIX_KMAP_BEGIN - 1)
 192                && vaddr > fix_to_virt(FIX_KMAP_END)
 193                && lastpte && lastpte + PTRS_PER_PTE != pte);
 194 #endif
 195         return pte;
 196 }
 197
 198 /*
 199  * This function initializes a certain range of kernel virtual memory
 200  * with new bootmem page tables, everywhere page tables are missing in
 201  * the given range.
 202  *
 203  * NOTE: The pagetables are allocated contiguous on the physical space
 204  * so we can cache the place of the first one and move around without
 205  * checking the pgd every time.
 206  */
 207 static void __init
 208 page_table_range_init(unsigned long start, unsigned long end, pgd_t *pgd_base)
 209 {
 210         int pgd_idx, pmd_idx;
 211         unsigned long vaddr;
 212         pgd_t *pgd;
 213         pmd_t *pmd;
 214         pte_t *pte = NULL;
 215         unsigned long count = page_table_range_init_count(start, end);
 216         void *adr = NULL;
 217
 218         if (count)
 219                 adr = alloc_low_pages(count);
 220
 221         vaddr = start;
 222         pgd_idx = pgd_index(vaddr);
 223         pmd_idx = pmd_index(vaddr);
 224         pgd = pgd_base + pgd_idx;
 225
 226         for ( ; (pgd_idx < PTRS_PER_PGD) && (vaddr != end); pgd++, pgd_idx++) {
 227                 pmd = one_md_table_init(pgd);
 228                 pmd = pmd + pmd_index(vaddr);
 229                 for (; (pmd_idx < PTRS_PER_PMD) && (vaddr != end);
 230                                                         pmd++, pmd_idx++) {
 231                         pte = page_table_kmap_check(one_page_table_init(pmd),
 232                                                     pmd, vaddr, pte, &adr);
 233
 234                         vaddr += PMD_SIZE;
 235                 }
 236                 pmd_idx = 0;
 237         }
 238 }
 239
 240 static inline int is_kernel_text(unsigned long addr)
 241 {
 242         if (addr >= (unsigned long)_text && addr <= (unsigned long)__init_end)
 243                 return 1;
 244         return 0;
 245 }
 246
 247 /*
 248  * This maps the physical memory to kernel virtual address space, a total
 249  * of max_low_pfn pages, by creating page tables starting from address
 250  * PAGE_OFFSET:
 251  */
 252 unsigned long __init
 253 kernel_physical_mapping_init(unsigned long start,
 254                              unsigned long end,
 255                              unsigned long page_size_mask)
 256 {
 257         int use_pse = page_size_mask == (1<<PG_LEVEL_2M);
 258         unsigned long last_map_addr = end;
 259         unsigned long start_pfn, end_pfn;
 260         pgd_t *pgd_base = swapper_pg_dir;
 261         int pgd_idx, pmd_idx, pte_ofs;
 262         unsigned long pfn;
 263         pgd_t *pgd;
 264         pmd_t *pmd;
 265         pte_t *pte;
 266         unsigned pages_2m, pages_4k;
 267         int mapping_iter;
 268
 269         start_pfn = start >> PAGE_SHIFT;
 270         end_pfn = end >> PAGE_SHIFT;
 271
 272         /*
 273          * First iteration will setup identity mapping using large/small pages
 274          * based on use_pse, with other attributes same as set by
 275          * the early code in head_32.S
 276          *
 277          * Second iteration will setup the appropriate attributes (NX, GLOBAL..)
 278          * as desired for the kernel identity mapping.
 279          *
 280          * This two pass mechanism conforms to the TLB app note which says:
 281          *
 282          *     "Software should not write to a paging-structure entry in a way
 283          *      that would change, for any linear address, both the page size
 284          *      and either the page frame or attributes."
 285          */
 286         mapping_iter = 1;
 287
 288         if (!boot_cpu_has(X86_FEATURE_PSE))
 289                 use_pse = 0;
 290
 291 repeat:
 292         pages_2m = pages_4k = 0;
 293         pfn = start_pfn;
 294         pgd_idx = pgd_index((pfn<<PAGE_SHIFT) + PAGE_OFFSET);
 295         pgd = pgd_base + pgd_idx;
 296         for (; pgd_idx < PTRS_PER_PGD; pgd++, pgd_idx++) {
 297                 pmd = one_md_table_init(pgd);
 298
 299                 if (pfn >= end_pfn)
 300                         continue;
 301 #ifdef CONFIG_X86_PAE
 302                 pmd_idx = pmd_index((pfn<<PAGE_SHIFT) + PAGE_OFFSET);
 303                 pmd += pmd_idx;
 304 #else
 305                 pmd_idx = 0;
 306 #endif
 307                 for (; pmd_idx < PTRS_PER_PMD && pfn < end_pfn;
 308                      pmd++, pmd_idx++) {
 309                         unsigned int addr = pfn * PAGE_SIZE + PAGE_OFFSET;
 310
 311                         /*
 312                          * Map with big pages if possible, otherwise
 313                          * create normal page tables:
 314                          */
 315                         if (use_pse) {
 316                                 unsigned int addr2;
 317                                 pgprot_t prot = PAGE_KERNEL_LARGE;
 318                                 /*
 319                                  * first pass will use the same initial
 320                                  * identity mapping attribute + _PAGE_PSE.
 321                                  */
 322                                 pgprot_t init_prot =
 323                                         __pgprot(PTE_IDENT_ATTR |
 324                                                  _PAGE_PSE);
 325
 326                                 pfn &= PMD_MASK >> PAGE_SHIFT;
 327                                 addr2 = (pfn + PTRS_PER_PTE-1) * PAGE_SIZE +
 328                                         PAGE_OFFSET + PAGE_SIZE-1;
 329
 330                                 if (is_kernel_text(addr) ||
 331                                     is_kernel_text(addr2))
 332                                         prot = PAGE_KERNEL_LARGE_EXEC;
 333
 334                                 pages_2m++;
 335                                 if (mapping_iter == 1)
 336                                         set_pmd(pmd, pfn_pmd(pfn, init_prot));
 337                                 else
 338                                         set_pmd(pmd, pfn_pmd(pfn, prot));
 339
 340                                 pfn += PTRS_PER_PTE;
 341                                 continue;
 342                         }
 343                         pte = one_page_table_init(pmd);
 344
 345                         pte_ofs = pte_index((pfn<<PAGE_SHIFT) + PAGE_OFFSET);
 346                         pte += pte_ofs;
 347                         for (; pte_ofs < PTRS_PER_PTE && pfn < end_pfn;
 348                              pte++, pfn++, pte_ofs++, addr += PAGE_SIZE) {
 349                                 pgprot_t prot = PAGE_KERNEL;
 350                                 /*
 351                                  * first pass will use the same initial
 352                                  * identity mapping attribute.
 353                                  */
 354                                 pgprot_t init_prot = __pgprot(PTE_IDENT_ATTR);
 355
 356                                 if (is_kernel_text(addr))
 357                                         prot = PAGE_KERNEL_EXEC;
 358
 359                                 pages_4k++;
 360                                 if (mapping_iter == 1) {
 361                                         set_pte(pte, pfn_pte(pfn, init_prot));
 362                                         last_map_addr = (pfn << PAGE_SHIFT) + PAGE_SIZE;
 363                                 } else
 364                                         set_pte(pte, pfn_pte(pfn, prot));
 365                         }
 366                 }
 367         }
 368         if (mapping_iter == 1) {
 369                 /*
 370                  * update direct mapping page count only in the first
 371                  * iteration.
 372                  */
 373                 update_page_count(PG_LEVEL_2M, pages_2m);
 374                 update_page_count(PG_LEVEL_4K, pages_4k);
 375
 376                 /*
 377                  * local global flush tlb, which will flush the previous
 378                  * mappings present in both small and large page TLB's.
 379                  */
 380                 __flush_tlb_all();
 381
 382                 /*
 383                  * Second iteration will set the actual desired PTE attributes.
 384                  */
 385                 mapping_iter = 2;
 386                 goto repeat;
 387         }
 388         return last_map_addr;
 389 }
 390
 391 pte_t *kmap_pte;
 392
 393 static inline pte_t *kmap_get_fixmap_pte(unsigned long vaddr)
 394 {
 395         pgd_t *pgd = pgd_offset_k(vaddr);
 396         p4d_t *p4d = p4d_offset(pgd, vaddr);
 397         pud_t *pud = pud_offset(p4d, vaddr);
 398         pmd_t *pmd = pmd_offset(pud, vaddr);
 399         return pte_offset_kernel(pmd, vaddr);
 400 }
 401
 402 static void __init kmap_init(void)
 403 {
 404         unsigned long kmap_vstart;
 405
 406         /*
 407          * Cache the first kmap pte:
 408          */
 409         kmap_vstart = __fix_to_virt(FIX_KMAP_BEGIN);
 410         kmap_pte = kmap_get_fixmap_pte(kmap_vstart);
 411 }
 412
 413 #ifdef CONFIG_HIGHMEM
 414 static void __init permanent_kmaps_init(pgd_t *pgd_base)
 415 {
 416         unsigned long vaddr;
 417         pgd_t *pgd;
 418         p4d_t *p4d;
 419         pud_t *pud;
 420         pmd_t *pmd;
 421         pte_t *pte;
 422
 423         vaddr = PKMAP_BASE;
 424         page_table_range_init(vaddr, vaddr + PAGE_SIZE*LAST_PKMAP, pgd_base);
 425
 426         pgd = swapper_pg_dir + pgd_index(vaddr);
 427         p4d = p4d_offset(pgd, vaddr);
 428         pud = pud_offset(p4d, vaddr);
 429         pmd = pmd_offset(pud, vaddr);
 430         pte = pte_offset_kernel(pmd, vaddr);
 431         pkmap_page_table = pte;
 432 }
 433
 434 void __init add_highpages_with_active_regions(int nid,
 435                          unsigned long start_pfn, unsigned long end_pfn)
 436 {
 437         phys_addr_t start, end;
 438         u64 i;
 439
 440         for_each_free_mem_range(i, nid, MEMBLOCK_NONE, &start, &end, NULL) {
 441                 unsigned long pfn = clamp_t(unsigned long, PFN_UP(start),
 442                                             start_pfn, end_pfn);
 443                 unsigned long e_pfn = clamp_t(unsigned long, PFN_DOWN(end),
 444                                               start_pfn, end_pfn);
 445                 for ( ; pfn < e_pfn; pfn++)
 446                         if (pfn_valid(pfn))
 447                                 free_highmem_page(pfn_to_page(pfn));
 448         }
 449 }
 450 #else
 451 static inline void permanent_kmaps_init(pgd_t *pgd_base)
 452 {
 453 }
 454 #endif /* CONFIG_HIGHMEM */
 455
 456 void __init native_pagetable_init(void)
 457 {
 458         unsigned long pfn, va;
 459         pgd_t *pgd, *base = swapper_pg_dir;
 460         p4d_t *p4d;
 461         pud_t *pud;
 462         pmd_t *pmd;
 463         pte_t *pte;
 464
 465         /*
 466          * Remove any mappings which extend past the end of physical
 467          * memory from the boot time page table.
 468          * In virtual address space, we should have at least two pages
 469          * from VMALLOC_END to pkmap or fixmap according to VMALLOC_END
 470          * definition. And max_low_pfn is set to VMALLOC_END physical
 471          * address. If initial memory mapping is doing right job, we
 472          * should have pte used near max_low_pfn or one pmd is not present.
 473          */
 474         for (pfn = max_low_pfn; pfn < 1<<(32-PAGE_SHIFT); pfn++) {
 475                 va = PAGE_OFFSET + (pfn<<PAGE_SHIFT);
 476                 pgd = base + pgd_index(va);
 477                 if (!pgd_present(*pgd))
 478                         break;
 479
 480                 p4d = p4d_offset(pgd, va);
 481                 pud = pud_offset(p4d, va);
 482                 pmd = pmd_offset(pud, va);
 483                 if (!pmd_present(*pmd))
 484                         break;
 485
 486                 /* should not be large page here */
 487                 if (pmd_large(*pmd)) {
 488                         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",
 489                                 pfn, pmd, __pa(pmd));
 490                         BUG_ON(1);
 491                 }
 492
 493                 pte = pte_offset_kernel(pmd, va);
 494                 if (!pte_present(*pte))
 495                         break;
 496
 497                 printk(KERN_DEBUG "clearing pte for ram above max_low_pfn: pfn: %lx pmd: %p pmd phys: %lx pte: %p pte phys: %lx\n",
 498                                 pfn, pmd, __pa(pmd), pte, __pa(pte));
 499                 pte_clear(NULL, va, pte);
 500         }
 501         paravirt_alloc_pmd(&init_mm, __pa(base) >> PAGE_SHIFT);
 502         paging_init();
 503 }
 504
 505 /*
 506  * Build a proper pagetable for the kernel mappings.  Up until this
 507  * point, we've been running on some set of pagetables constructed by
 508  * the boot process.
 509  *
 510  * If we're booting on native hardware, this will be a pagetable
 511  * constructed in arch/x86/kernel/head_32.S.  The root of the
 512  * pagetable will be swapper_pg_dir.
 513  *
 514  * If we're booting paravirtualized under a hypervisor, then there are
 515  * more options: we may already be running PAE, and the pagetable may
 516  * or may not be based in swapper_pg_dir.  In any case,
 517  * paravirt_pagetable_init() will set up swapper_pg_dir
 518  * appropriately for the rest of the initialization to work.
 519  *
 520  * In general, pagetable_init() assumes that the pagetable may already
 521  * be partially populated, and so it avoids stomping on any existing
 522  * mappings.
 523  */
 524 void __init early_ioremap_page_table_range_init(void)
 525 {
 526         pgd_t *pgd_base = swapper_pg_dir;
 527         unsigned long vaddr, end;
 528
 529         /*
 530          * Fixed mappings, only the page table structure has to be
 531          * created - mappings will be set by set_fixmap():
 532          */
 533         vaddr = __fix_to_virt(__end_of_fixed_addresses - 1) & PMD_MASK;
 534         end = (FIXADDR_TOP + PMD_SIZE - 1) & PMD_MASK;
 535         page_table_range_init(vaddr, end, pgd_base);
 536         early_ioremap_reset();
 537 }
 538
 539 static void __init pagetable_init(void)
 540 {
 541         pgd_t *pgd_base = swapper_pg_dir;
 542
 543         permanent_kmaps_init(pgd_base);
 544 }
 545
 546 pteval_t __supported_pte_mask __read_mostly = ~(_PAGE_NX | _PAGE_GLOBAL);
 547 EXPORT_SYMBOL_GPL(__supported_pte_mask);
 548
 549 /* user-defined highmem size */
 550 static unsigned int highmem_pages = -1;
 551
 552 /*
 553  * highmem=size forces highmem to be exactly 'size' bytes.
 554  * This works even on boxes that have no highmem otherwise.
 555  * This also works to reduce highmem size on bigger boxes.
 556  */
 557 static int __init parse_highmem(char *arg)
 558 {
 559         if (!arg)
 560                 return -EINVAL;
 561
 562         highmem_pages = memparse(arg, &arg) >> PAGE_SHIFT;
 563         return 0;
 564 }
 565 early_param("highmem", parse_highmem);
 566
 567 #define MSG_HIGHMEM_TOO_BIG \
 568         "highmem size (%luMB) is bigger than pages available (%luMB)!\n"
 569
 570 #define MSG_LOWMEM_TOO_SMALL \
 571         "highmem size (%luMB) results in <64MB lowmem, ignoring it!\n"
 572 /*
 573  * All of RAM fits into lowmem - but if user wants highmem
 574  * artificially via the highmem=x boot parameter then create
 575  * it:
 576  */
 577 static void __init lowmem_pfn_init(void)
 578 {
 579         /* max_low_pfn is 0, we already have early_res support */
 580         max_low_pfn = max_pfn;
 581
 582         if (highmem_pages == -1)
 583                 highmem_pages = 0;
 584 #ifdef CONFIG_HIGHMEM
 585         if (highmem_pages >= max_pfn) {
 586                 printk(KERN_ERR MSG_HIGHMEM_TOO_BIG,
 587                         pages_to_mb(highmem_pages), pages_to_mb(max_pfn));
 588                 highmem_pages = 0;
 589         }
 590         if (highmem_pages) {
 591                 if (max_low_pfn - highmem_pages < 64*1024*1024/PAGE_SIZE) {
 592                         printk(KERN_ERR MSG_LOWMEM_TOO_SMALL,
 593                                 pages_to_mb(highmem_pages));
 594                         highmem_pages = 0;
 595                 }
 596                 max_low_pfn -= highmem_pages;
 597         }
 598 #else
 599         if (highmem_pages)
 600                 printk(KERN_ERR "ignoring highmem size on non-highmem kernel!\n");
 601 #endif
 602 }
 603
 604 #define MSG_HIGHMEM_TOO_SMALL \
 605         "only %luMB highmem pages available, ignoring highmem size of %luMB!\n"
 606
 607 #define MSG_HIGHMEM_TRIMMED \
 608         "Warning: only 4GB will be used. Use a HIGHMEM64G enabled kernel!\n"
 609 /*
 610  * We have more RAM than fits into lowmem - we try to put it into
 611  * highmem, also taking the highmem=x boot parameter into account:
 612  */
 613 static void __init highmem_pfn_init(void)
 614 {
 615         max_low_pfn = MAXMEM_PFN;
 616
 617         if (highmem_pages == -1)
 618                 highmem_pages = max_pfn - MAXMEM_PFN;
 619
 620         if (highmem_pages + MAXMEM_PFN < max_pfn)
 621                 max_pfn = MAXMEM_PFN + highmem_pages;
 622
 623         if (highmem_pages + MAXMEM_PFN > max_pfn) {
 624                 printk(KERN_WARNING MSG_HIGHMEM_TOO_SMALL,
 625                         pages_to_mb(max_pfn - MAXMEM_PFN),
 626                         pages_to_mb(highmem_pages));
 627                 highmem_pages = 0;
 628         }
 629 #ifndef CONFIG_HIGHMEM
 630         /* Maximum memory usable is what is directly addressable */
 631         printk(KERN_WARNING "Warning only %ldMB will be used.\n", MAXMEM>>20);
 632         if (max_pfn > MAX_NONPAE_PFN)
 633                 printk(KERN_WARNING "Use a HIGHMEM64G enabled kernel.\n");
 634         else
 635                 printk(KERN_WARNING "Use a HIGHMEM enabled kernel.\n");
 636         max_pfn = MAXMEM_PFN;
 637 #else /* !CONFIG_HIGHMEM */
 638 #ifndef CONFIG_HIGHMEM64G
 639         if (max_pfn > MAX_NONPAE_PFN) {
 640                 max_pfn = MAX_NONPAE_PFN;
 641                 printk(KERN_WARNING MSG_HIGHMEM_TRIMMED);
 642         }
 643 #endif /* !CONFIG_HIGHMEM64G */
 644 #endif /* !CONFIG_HIGHMEM */
 645 }
 646
 647 /*
 648  * Determine low and high memory ranges:
 649  */
 650 void __init find_low_pfn_range(void)
 651 {
 652         /* it could update max_pfn */
 653
 654         if (max_pfn <= MAXMEM_PFN)
 655                 lowmem_pfn_init();
 656         else
 657                 highmem_pfn_init();
 658 }
 659
 660 #ifndef CONFIG_NEED_MULTIPLE_NODES
 661 void __init initmem_init(void)
 662 {
 663 #ifdef CONFIG_HIGHMEM
 664         highstart_pfn = highend_pfn = max_pfn;
 665         if (max_pfn > max_low_pfn)
 666                 highstart_pfn = max_low_pfn;
 667         printk(KERN_NOTICE "%ldMB HIGHMEM available.\n",
 668                 pages_to_mb(highend_pfn - highstart_pfn));
 669         high_memory = (void *) __va(highstart_pfn * PAGE_SIZE - 1) + 1;
 670 #else
 671         high_memory = (void *) __va(max_low_pfn * PAGE_SIZE - 1) + 1;
 672 #endif
 673
 674         memblock_set_node(0, (phys_addr_t)ULLONG_MAX, &memblock.memory, 0);
 675         sparse_memory_present_with_active_regions(0);
 676
 677 #ifdef CONFIG_FLATMEM
 678         max_mapnr = IS_ENABLED(CONFIG_HIGHMEM) ? highend_pfn : max_low_pfn;
 679 #endif
 680         __vmalloc_start_set = true;
 681
 682         printk(KERN_NOTICE "%ldMB LOWMEM available.\n",
 683                         pages_to_mb(max_low_pfn));
 684
 685         setup_bootmem_allocator();
 686 }
 687 #endif /* !CONFIG_NEED_MULTIPLE_NODES */
 688
 689 void __init setup_bootmem_allocator(void)
 690 {
 691         printk(KERN_INFO "  mapped low ram: 0 - %08lx\n",
 692                  max_pfn_mapped<<PAGE_SHIFT);
 693         printk(KERN_INFO "  low ram: 0 - %08lx\n", max_low_pfn<<PAGE_SHIFT);
 694 }
 695
 696 /*
 697  * paging_init() sets up the page tables - note that the first 8MB are
 698  * already mapped by head.S.
 699  *
 700  * This routines also unmaps the page at virtual kernel address 0, so
 701  * that we can trap those pesky NULL-reference errors in the kernel.
 702  */
 703 void __init paging_init(void)
 704 {
 705         pagetable_init();
 706
 707         __flush_tlb_all();
 708
 709         kmap_init();
 710
 711         /*
 712          * NOTE: at this point the bootmem allocator is fully available.
 713          */
 714         olpc_dt_build_devicetree();
 715         sparse_memory_present_with_active_regions(MAX_NUMNODES);
 716         sparse_init();
 717         zone_sizes_init();
 718 }
 719
 720 /*
 721  * Test if the WP bit works in supervisor mode. It isn't supported on 386's
 722  * and also on some strange 486's. All 586+'s are OK. This used to involve
 723  * black magic jumps to work around some nasty CPU bugs, but fortunately the
 724  * switch to using exceptions got rid of all that.
 725  */
 726 static void __init test_wp_bit(void)
 727 {
 728         char z = 0;
 729
 730         printk(KERN_INFO "Checking if this processor honours the WP bit even in supervisor mode...");
 731
 732         __set_fixmap(FIX_WP_TEST, __pa_symbol(empty_zero_page), PAGE_KERNEL_RO);
 733
 734         if (probe_kernel_write((char *)fix_to_virt(FIX_WP_TEST), &z, 1)) {
 735                 clear_fixmap(FIX_WP_TEST);
 736                 printk(KERN_CONT "Ok.\n");
 737                 return;
 738         }
 739
 740         printk(KERN_CONT "No.\n");
 741         panic("Linux doesn't support CPUs with broken WP.");
 742 }
 743
 744 void __init mem_init(void)
 745 {
 746         pci_iommu_alloc();
 747
 748 #ifdef CONFIG_FLATMEM
 749         BUG_ON(!mem_map);
 750 #endif
 751         /*
 752          * With CONFIG_DEBUG_PAGEALLOC initialization of highmem pages has to
 753          * be done before free_all_bootmem(). Memblock use free low memory for
 754          * temporary data (see find_range_array()) and for this purpose can use
 755          * pages that was already passed to the buddy allocator, hence marked as
 756          * not accessible in the page tables when compiled with
 757          * CONFIG_DEBUG_PAGEALLOC. Otherwise order of initialization is not
 758          * important here.
 759          */
 760         set_highmem_pages_init();
 761
 762         /* this will put all low memory onto the freelists */
 763         free_all_bootmem();
 764
 765         after_bootmem = 1;
 766
 767         mem_init_print_info(NULL);
 768         printk(KERN_INFO "virtual kernel memory layout:\n"
 769                 "    fixmap  : 0x%08lx - 0x%08lx   (%4ld kB)\n"
 770                 "  cpu_entry : 0x%08lx - 0x%08lx   (%4ld kB)\n"
 771 #ifdef CONFIG_HIGHMEM
 772                 "    pkmap   : 0x%08lx - 0x%08lx   (%4ld kB)\n"
 773 #endif
 774                 "    vmalloc : 0x%08lx - 0x%08lx   (%4ld MB)\n"
 775                 "    lowmem  : 0x%08lx - 0x%08lx   (%4ld MB)\n"
 776                 "      .init : 0x%08lx - 0x%08lx   (%4ld kB)\n"
 777                 "      .data : 0x%08lx - 0x%08lx   (%4ld kB)\n"
 778                 "      .text : 0x%08lx - 0x%08lx   (%4ld kB)\n",
 779                 FIXADDR_START, FIXADDR_TOP,
 780                 (FIXADDR_TOP - FIXADDR_START) >> 10,
 781
 782                 CPU_ENTRY_AREA_BASE,
 783                 CPU_ENTRY_AREA_BASE + CPU_ENTRY_AREA_MAP_SIZE,
 784                 CPU_ENTRY_AREA_MAP_SIZE >> 10,
 785
 786 #ifdef CONFIG_HIGHMEM
 787                 PKMAP_BASE, PKMAP_BASE+LAST_PKMAP*PAGE_SIZE,
 788                 (LAST_PKMAP*PAGE_SIZE) >> 10,
 789 #endif
 790
 791                 VMALLOC_START, VMALLOC_END,
 792                 (VMALLOC_END - VMALLOC_START) >> 20,
 793
 794                 (unsigned long)__va(0), (unsigned long)high_memory,
 795                 ((unsigned long)high_memory - (unsigned long)__va(0)) >> 20,
 796
 797                 (unsigned long)&__init_begin, (unsigned long)&__init_end,
 798                 ((unsigned long)&__init_end -
 799                  (unsigned long)&__init_begin) >> 10,
 800
 801                 (unsigned long)&_etext, (unsigned long)&_edata,
 802                 ((unsigned long)&_edata - (unsigned long)&_etext) >> 10,
 803
 804                 (unsigned long)&_text, (unsigned long)&_etext,
 805                 ((unsigned long)&_etext - (unsigned long)&_text) >> 10);
 806
 807         /*
 808          * Check boundaries twice: Some fundamental inconsistencies can
 809          * be detected at build time already.
 810          */
 811 #define __FIXADDR_TOP (-PAGE_SIZE)
 812 #ifdef CONFIG_HIGHMEM
 813         BUILD_BUG_ON(PKMAP_BASE + LAST_PKMAP*PAGE_SIZE  > FIXADDR_START);
 814         BUILD_BUG_ON(VMALLOC_END                        > PKMAP_BASE);
 815 #endif
 816 #define high_memory (-128UL << 20)
 817         BUILD_BUG_ON(VMALLOC_START                      >= VMALLOC_END);
 818 #undef high_memory
 819 #undef __FIXADDR_TOP
 820
 821 #ifdef CONFIG_HIGHMEM
 822         BUG_ON(PKMAP_BASE + LAST_PKMAP*PAGE_SIZE        > FIXADDR_START);
 823         BUG_ON(VMALLOC_END                              > PKMAP_BASE);
 824 #endif
 825         BUG_ON(VMALLOC_START                            >= VMALLOC_END);
 826         BUG_ON((unsigned long)high_memory               > VMALLOC_START);
 827
 828         test_wp_bit();
 829 }
 830
 831 #ifdef CONFIG_MEMORY_HOTPLUG
 832 int arch_add_memory(int nid, u64 start, u64 size, bool want_memblock)
 833 {
 834         unsigned long start_pfn = start >> PAGE_SHIFT;
 835         unsigned long nr_pages = size >> PAGE_SHIFT;
 836
 837         return __add_pages(nid, start_pfn, nr_pages, want_memblock);
 838 }
 839
 840 #ifdef CONFIG_MEMORY_HOTREMOVE
 841 int arch_remove_memory(u64 start, u64 size)
 842 {
 843         unsigned long start_pfn = start >> PAGE_SHIFT;
 844         unsigned long nr_pages = size >> PAGE_SHIFT;
 845         struct zone *zone;
 846
 847         zone = page_zone(pfn_to_page(start_pfn));
 848         return __remove_pages(zone, start_pfn, nr_pages);
 849 }
 850 #endif
 851 #endif
 852
 853 int kernel_set_to_readonly __read_mostly;
 854
 855 void set_kernel_text_rw(void)
 856 {
 857         unsigned long start = PFN_ALIGN(_text);
 858         unsigned long size = PFN_ALIGN(_etext) - start;
 859
 860         if (!kernel_set_to_readonly)
 861                 return;
 862
 863         pr_debug("Set kernel text: %lx - %lx for read write\n",
 864                  start, start+size);
 865
 866         set_pages_rw(virt_to_page(start), size >> PAGE_SHIFT);
 867 }
 868
 869 void set_kernel_text_ro(void)
 870 {
 871         unsigned long start = PFN_ALIGN(_text);
 872         unsigned long size = PFN_ALIGN(_etext) - start;
 873
 874         if (!kernel_set_to_readonly)
 875                 return;
 876
 877         pr_debug("Set kernel text: %lx - %lx for read only\n",
 878                  start, start+size);
 879
 880         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
 881 }
 882
 883 static void mark_nxdata_nx(void)
 884 {
 885         /*
 886          * When this called, init has already been executed and released,
 887          * so everything past _etext should be NX.
 888          */
 889         unsigned long start = PFN_ALIGN(_etext);
 890         /*
 891          * This comes from is_kernel_text upper limit. Also HPAGE where used:
 892          */
 893         unsigned long size = (((unsigned long)__init_end + HPAGE_SIZE) & HPAGE_MASK) - start;
 894
 895         if (__supported_pte_mask & _PAGE_NX)
 896                 printk(KERN_INFO "NX-protecting the kernel data: %luk\n", size >> 10);
 897         set_pages_nx(virt_to_page(start), size >> PAGE_SHIFT);
 898 }
 899
 900 void mark_rodata_ro(void)
 901 {
 902         unsigned long start = PFN_ALIGN(_text);
 903         unsigned long size = PFN_ALIGN(_etext) - start;
 904
 905         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
 906         printk(KERN_INFO "Write protecting the kernel text: %luk\n",
 907                 size >> 10);
 908
 909         kernel_set_to_readonly = 1;
 910
 911 #ifdef CONFIG_CPA_DEBUG
 912         printk(KERN_INFO "Testing CPA: Reverting %lx-%lx\n",
 913                 start, start+size);
 914         set_pages_rw(virt_to_page(start), size>>PAGE_SHIFT);
 915
 916         printk(KERN_INFO "Testing CPA: write protecting again\n");
 917         set_pages_ro(virt_to_page(start), size>>PAGE_SHIFT);
 918 #endif
 919
 920         start += size;
 921         size = (unsigned long)__end_rodata - start;
 922         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
 923         printk(KERN_INFO "Write protecting the kernel read-only data: %luk\n",
 924                 size >> 10);
 925
 926 #ifdef CONFIG_CPA_DEBUG
 927         printk(KERN_INFO "Testing CPA: undo %lx-%lx\n", start, start + size);
 928         set_pages_rw(virt_to_page(start), size >> PAGE_SHIFT);
 929
 930         printk(KERN_INFO "Testing CPA: write protecting again\n");
 931         set_pages_ro(virt_to_page(start), size >> PAGE_SHIFT);
 932 #endif
 933         mark_nxdata_nx();
 934         if (__supported_pte_mask & _PAGE_NX)
 935                 debug_checkwx();
 936 }